The present application is concerned with an improved intra-prediction mode concept for block-wise picture coding such as usable in a video codec such as HEVC or any succes-sor of HEVC.

Intra-prediction modes are widely used in picture and video coding. In video coding, intra-prediction modes compete with other prediction modes such as inter-prediction modes such as motion-compensated prediction modes. In intra-prediction modes, a current block is predicted on the basis of neighboring samples, i.e. samples already encoded as far as the encoder side is concerned, and already decoded as far as the decoder side is con-cerned. Neighboring sample values are extrapolated into the current block so as to form a prediction signal for the current block with the prediction residual being transmitted in the datastream for the current block. The better the prediction signal is, the lower the predic-tion residual is and, accordingly, a lower number of bits is necessary to code the predic-tion residual.

In order to be effective, several aspects should be taken into account in order to form an effective frame work for intra-prediction in a block-wise picture coding environment. For instance, the larger the number of intra-prediction modes supported by the codec, the larger the side information rate consumption is in order to signal the selection to the de-coder. On the other hand, the set of supported intra-prediction modes should be able to provide a good prediction signal, i.e. a prediction signal resulting in a low prediction resid-ual.

The present application seeks to provide an intra-prediction mode concept allowing for a more efficient compression of a block-wise picture codec if using the improved intra-prediction mode concept.

This object is achieved by the subject-matter of the independent claims of the present application.

Summary

There is disclosed an apparatus (e.g., decoder) for block-wise decoding a picture from a data stream, the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal for a block of a predetermined size of the picture is de-termined by applying a first template of samples which neighbours the current block onto a neural network, wherein the apparatus is configured, for a current block differing from the predetermined size, to:

resample a second template of samples neighboring the current block, so as to conform with the first template so as to obtain a resampled template;

apply the resampled template of samples onto the neural network so as to obtain a preliminary intra-prediction; and

resample the preliminary intra-prediction signal so as to conform with the current block so as to obtain the intra-prediction signal for the current block.

There is also disclosed an apparatus (e.g., encoder) for block-wise encoding a picture into a data stream, the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal for a block of a predetermined size of the picture is de-termined by applying a first template of samples which neighbours the current block onto a neural network, wherein the apparatus is configured, for a current block differing from the predetermined size, to:

resample a second template of samples neighboring the current block, so as to conform with the first template so as to obtain a resampled template;

apply the resampled template of samples onto the neural network so as to obtain a preliminary intra-prediction; and

resample the preliminary intra-prediction signal so as to conform with the current block so as to obtain the intra-prediction signal for the current block.

The apparatus may be configured to resample by downsampling the second template to obtain the first template.

The apparatus may be configured to resample the preliminary intra-prediction signal by upsampling the preliminary intra-prediction signal.

The apparatus may be configured to transform the preliminary intra-prediction signal from a spatial domain into a transform domain; and resample the preliminary intra-prediction signal in the transform domain.

The apparatus may be configured to resample the transform-domain preliminary intra-prediction signal by scaling the coefficients of the preliminary intra-prediction signal.

The apparatus may be configured to

resample the transform-domain preliminary intra-prediction signal by:

increasing the dimensions of the intra-prediction signal to conform to the dimensions of the current block; and

zero-padding the coefficients of added coefficients of the preliminary intra- prediction signal, the added coefficients relating to higher-frequencies bins.

The apparatus may be configured to compose the transform-domain preliminary intra-prediction signal with a dequantized version of a prediction residual signal.

The apparatus may be configured to resample the preliminary intra-prediction signal in the spatial domain.

The apparatus may be configured to resample the preliminary intra-prediction signal by performing a bilinear interpolation.

The apparatus may be configured to encode in a data field information regarding the resampling and/or the use of neural networks for different dimensions.

There is also disclosed an apparatus (e.g., decoder) for block-wise decoding a picture from a data stream, the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal for a current block of the picture is determined by: applying a first set of neighboring samples of the current block onto a neural net-work to obtain a prediction of a set of transform coefficients of a transform of the current block.

There is also disclosed an apparatus (e.g., encoder) for block-wise encoding a picture into a data stream, the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal for a current block of the picture is determined by:

applying a first set of neighboring samples of the current block onto a neural net-work to obtain a prediction of a set of transform coefficients of a transform of the current block.

One of the apparatus may be configured to inversely transform the prediction to obtain a reconstructed signal.

One of the apparatus may be configured to decode from the data stream an index using a variable length code; and perform the selection using the index.

One of the apparatus may be configured to determine a ranking of the set of intra predic-tion modes; and, subsequently, resample the second template.

There is disclosed a method comprising:

resampling a second template of samples neighboring the current block, so as to conform with a first template so as to obtain a resampled template,

applying the resampled template of samples onto a neural network so as to obtain a preliminary intra-prediction signal, and

resampling the preliminary intra-prediction signal so as to conform with the current block so as to obtain the intra-prediction signal for the current block.

There is disclosed a method for block-wise decoding a picture from a data stream, com-prising:

applying a first set of neighboring samples of a current block onto a neural network to obtain a prediction of a set of transform coefficients of a transform of a current block.

There is disclosed a method for block-wise encoding a picture into a data stream, corn-prising:

applying a first set of neighboring samples of a current block onto a neural network to obtain a prediction of a set of transform coefficients of a transform of a current block.

A method of above and/or below may use the equipment comprising at least one appa-ratus as above and/or below.

There is also disclosed a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to perform a method as above and/or below and/or implementing at least one component of the apparatus above and/or below.

There is also disclosed a data stream obtained by a method as above and/or below and/or by an apparatus as above and/or below.

Figures

As far as the design of the above-mentioned neural networks is concerned, the present application provides many examples for appropriately determining parameters thereof.

Advantageous implementations of the present application are the subject of dependent claims. Preferred examples of the present application are described in the following with respect to the Figures, among which:

Fig. 1 shows a schematic block diagram illustrating an encoder for encoding a picture into a datastream as a general example where examples of the present application may be implemented;

Fig. 2 shows a block diagram of a more specific example for an encoder according to Fig.

1;

Fig. 3 shows a schematic block diagram illustrating a decoder fitting to the encoder of Fig.

1 and serving as an example for a decoder where examples of the present applica- tion may be implemented into;

Fig. 4 shows a block diagram of a more specific example of a decoder of Fig. 3, which fits to the encoder of Fig. 2;

Fig. 5 shows a schematic diagram illustrating the mode of operation in encoder and de- coder according to examples of the present application with respect to processing a block using intra-prediction;

Fig. 6 shows a schematic block diagram illustrating a decoder in accordance with an ex- ample of the present application comprising several neural network-based intra- prediction modes;

Fig. 7a shows a schematic diagram illustrating the mode of operation of an encoder and a decoder in accordance with an example supporting neural network-based intra- prediction modes and a neural network-based ordering of these modes with

transmitting within the datastream an index into an ordered list of neural network- based intra-prediction modes along with a fleck indicating whether the intra- prediction mode to be used is member of the set of neural network-based intra- prediction modes or not. It is needless to say the index may be coded using varia- ble length coding so as to take advantage of the different frequency of being de- termined by determination 90;

Fig. 7b shows a schematic diagram differing from Fig. 7a in that the fleck signalization is not used;

Fig. 7c shows a schematic diagram differing from 7b in that the mode ordering is not con- trolled using a neural network;

Fig. 7d shows a schematic diagram differing from 7a in that the neural network assistance in mode signalization is used for controlling the probability distribution estimation in entropy en/decoding rather than for mode ordering;

Fig. 8 shows an apparatus for designing a set of neural network-based intra-prediction modes in accordance with an example;

Fig. 9a shows a schematic diagram illustrating the mode of operation of encoder and de- coder in accordance with an example according to which a neural network is used for ordering supported intra-prediction modes irrespective of whether neural net- work-based or not;

Fig. 9b shows a schematic diagram differing from Fig. 9a in that the neural network base is used for controlling the probabilities distribution estimation for entropy de/encoding of the index into the set of supported intra-prediction modes;

Fig. 10 shows an apparatus for designing a neural network for assisting and selecting among a set of intra-prediction mode for block-based picture coding in accordance with an example.

Fig. 11-1 shows an encoder according to an example.

Fig. 11-2 shows a decoder according to an example.

Fig. 12 shows a schematic diagram illustrating the mode of operation of an encoder and a decoder in accordance with an example.

Figs. 13a and 13b show schematic diagrams of techniques according to examples.

Examples

In the following, various examples are described which assist in achieving a more effec-tive compression when using intra-prediction. Some examples achieve the compression efficiency increase by spending a set of intra-prediction modes which are neural network-based. The latter ones may be added to other intra-prediction modes heuristically de-signed, for instance, or may be provided exclusively. Other examples use a neural net-work in order to perform a selection among a plurality of intra-prediction modes. And even other examples make use of both of the just-discussed specialties.

In order to ease the understanding of the following examples of the present application, the description starts with a presentation of possible encoders and decoders fitting thereto into which the subsequently outlined examples of the present application could be built. Fig. 1 shows an apparatus for block-wise encoding a picture 10 into a datastream 12. The apparatus is indicated using reference sign 14 and may be a still picture encoder or a vid-eo encoder. In other words, picture 10 may be a current picture out of a video 16 when the encoder 14 is configured to encode video 16 including picture 10 into datastream 12, or encoder 14 may encode picture 10 into datastream 12 exclusively.

As mentioned, encoder 14 performs the encoding in a block-wise manner or block-base. To this, encoder 14 subdivides picture 10 into blocks, units of which encoder 14 encodes picture 10 into datastream 12. Examples of possible subdivisions of picture 10 into blocks 18 are set out in more detail below. Generally, the subdivision may end-up into blocks 18 of constant size such as an array of blocks arranged in rows and columns or into blocks 18 of different block sizes such as by use of a hierarchical multi-tree subdivisioning with starting the multi-tree subdivisioning from the whole picture area of picture 10 or from a pre-partitioning of picture 10 into an array of tree blocks wherein these examples shall not be treated as excluding other possible ways of subdivisioning picture 10 into blocks 18.

Further, encoder 14 is a predictive encoder configured to predictively encode picture 10 into datastream 12. For a certain block 18 this means that encoder 14 determines a pre-diction signal for block 18 and encodes the prediction residual, i.e. the prediction error at which the prediction signal deviates from the actual picture content within block 18, into datastream 12.

Encoder 14 may support different prediction modes so as to derive the prediction signal for a certain block 18. The prediction modes, which are of importance in the following ex-amples, are intra-prediction modes according to which the inner of block 18 is predicted spatially from neighboring, already encoded samples of picture 10. The encoding of pic-ture 10 into datastream 12 and, accordingly, the corresponding decoding procedure, may be based on a certain coding order 20 defined among blocks 18. For instance, the coding order 20 may traverse blocks 18 in a raster scan order such as row-wise from top to bot-tom with traversing each row from left to right, for instance. In case of hierarchical multi-tree based subdivisioning, raster scan ordering may be applied within each hierarchy lev-el, wherein a depth-first traversal order may be applied, i.e. leaf notes within a block of a certain hierarchy level may precede blocks of the same hierarchy level having the same parent block according to coding order 20. Depending on the coding order 20, neighbor-ing, already encoded samples of a block 18 may be located usually at one or more sides of block 18. In case of the examples presented herein, for instance, neighboring, already encoded samples of a block 18 are located to the top of, and to the left of block 18.

Intra-prediction modes may not be the only ones supported by encoder 14. In case of en-coder 14 being a video encoder, for instance, encoder 14 may also support intra-prediction modes according to which a block 18 is temporarily predicted from a previously encoded picture of video 16. Such an intra-prediction mode may be a motion-compensated prediction mode according to which a motion vector is signaled for such a block 18 indicating a relative spatial offset of the portion from which the prediction signal of block 18 is to be derived as a copy. Additionally or alternatively, other non-intra-prediction modes may be available as well such as inter-view prediction modes in case of encoder 14 being a multi-view encoder, or non-predictive modes according to which the inner of block 18 is coded as is, i.e. without any prediction.

Before starting with focusing the description of the present application onto intra-prediction modes, a more specific example for a possible block-based encoder, i.e. for a possible implementation of encoder 14, as described with respect to Fig. 2 with then presenting two corresponding examples for a decoder fitting to Figs. 1 and 2, respectively.

Fig. 2 shows a possible implementation of encoder 14 of Fig. 1, namely one where the encoder is configured to use transform coding for encoding the prediction residual alt-hough this is nearly an example and the present application is not restricted to that sort of prediction residual coding. According to Fig. 2, encoder 14 comprises a subtractor 22 con-figured to subtract from the inbound signal, i.e. picture 10 or, on a block basis, current block 18, the corresponding prediction signal 24 so as to obtain the prediction residual signal 26 which is then encoded by a prediction residual encoder 28 into a datastream 12. The prediction residual encoder 28 is composed of a lossy encoding stage 28a and a lossless encoding stage 28b. The lossy stage 28a receives the prediction residual signal 26 and comprises a quantizer 30 which quantizes the samples of the prediction residual signal 26. As already mentioned above, the present example uses transform coding of the prediction residual signal 26 and accordingly, the lossy encoding stage 28a comprises a transform stage 32 connected between subtractor 22 and quantizer 30 so as to transform such a spectrally decomposed prediction residual 26 with a quantization of quantizer 30 taking place on the transformed coefficients where presenting the residual signal 26. The transform may be a DCT, DST, FFT, Hadamard transform or the like. The transformed and quantized prediction residual signal 34 is then subject to lossless coding by the loss-less encoding stage 28b which is an entropy coder entropy coding quantized prediction residual signal 34 into datastream 12. Encoder 14 further comprises the prediction residu-al signal reconstruction stage 36 connected to the output of quantizer 30 so as to recon-struct from the transformed and quantized prediction residual signal 34 the prediction re-sidual signal in a manner also available at the decoder, i.e. taking the coding loss is quan-tizer 30 into account. To this end, the prediction residual reconstruction stage 36 compris-es a dequantizer 38 which perform the inverse of the quantization of quantizer 30, fol-lowed by an inverse transformer 40 which performs the inverse transformation relative to the transformation performed by transformer 32 such as the inverse of the spectral de-composition such as the inverse to any of the above-mentioned specific transformation examples. Encoder 14 comprises an adder 42 which adds the reconstructed prediction residual signal as output by inverse transformer 40 and the prediction signal 24 so as to output a reconstructed signal, i.e. reconstructed samples. This output is fed into a predic-tor 44 of encoder 14 which then determines the prediction signal 24 based thereon. It is predictor 44 which supports all the prediction modes already discussed above with re-spect to Fig. 1. Fig. 2 also illustrates that in case of encoder 14 being a video encoder, encoder 14 may also comprise an in-loop filter 46 with filters completely reconstructed pictures which, after having been filtered, form reference pictures for predictor 44 with respect to inter-predicted block.

As already mentioned above, encoder 14 operates block-based. For the subsequent de-scription, the block bases of interest is the one subdividing picture 10 into blocks for which the intra-prediction mode is selected out of a set or plurality of intra-prediction modes sup-ported by predictor 44 or encoder 14 respectively, and the selected intra-prediction mode performed individually. Other sorts of blocks into which picture 10 is subdivided may, however, exist as well. For instance, the above-mentioned decision whether picture 10 is inter-coded or intra-coded may be done at a granularity or in units of blocks deviating from blocks 18. For instance, the inter/intra mode decision may be performed at a level of cod-ing blocks into which picture 10 is subdivided, and each coding block is subdivided into prediction blocks. Prediction blocks with encoding blocks for which it has been decided that intra-prediction is used, are each subdivided to an intra-prediction mode decision. To this, for each of these prediction blocks, it is decided as to which supported intra-prediction mode should be used for the respective prediction block. These prediction blocks will form blocks 18 which are of interest here. Prediction blocks within coding blocks associated with inter-prediction would be treated differently by predictor 44. They would be inter-predicted from reference pictures by determining a motion vector and copy-ing the prediction signal for this block from a location in the reference picture pointed to by the motion vector. Another block subdivisioning pertains the subdivisioning into transform blocks at units of which the transformations by transformer 32 and inverse transformer 40 are performed. Transformed blocks may, for instance, be the result of further subdivision-ing coding blocks. Naturally, the examples set out herein should not be treated as being limiting and other examples exist as well. For the sake of completeness only, it is noted that the subdivisioning into coding blocks may, for instance, use multi-tree subdivisioning, and prediction blocks and/or transform blocks may be obtained by further subdividing cod-ing blocks using multi-tree subdivisioning, as well.

A decoder or apparatus for block-wise decoding fitting to the encoder 14 of Fig. 1 is de-picted in Fig. 3. This decoder 54 does the opposite of encoder 14, i.e. it decodes from datastream 12 picture 10 in a block-wise manner and supports, to this end, a plurality of intra-prediction modes. The decoder 54 may comprise a residual provider 156, for exam-ple. All the other possibilities discussed above with respect to Fig. 1 are valid for the de-coder 54, too. To this, decoder 54 may be a still picture decoder or a video decoder and all the prediction modes and prediction possibilities are supported by decoder 54 as well. The difference between encoder 14 and decoder 54 lies, primarily, in the fact that encoder 14 chooses or selects coding decisions according to some optimization such as, for in-

stance, in order to minimize some cost function which may depend on coding rate and/or coding distortion. One of these coding options or coding parameters may involve a selec-tion of the intra-prediction mode to be used for a current block 18 among available or sup-ported intra-prediction modes. The selected intra-prediction mode may then be signaled by encoder 14 for current block 18 within datastream 12 with decoder 54 redoing the se-lection using this signalization in datastream 12 for block 18. Likewise, the subdivisioning of picture 10 into blocks 18 may be subject to optimization within encoder 14 and corre-sponding subdivision information may be conveyed within datastream 12 with decoder 54 recovering the subdivision of picture 10 into blocks 18 on the basis of the subdivision in-formation. Summarizing the above, decoder 54 may be a predictive decoder operating on a block-bases and besides intra-prediction modes, decoder 54 may support other predic-tion modes such as inter-prediction modes in case of, for instance, decoder 54 being a video decoder. In decoding, decoder 54 may also use the coding order 20 discussed with respect to Fig. 1 and as this coding order 20 is obeyed both at encoder 14 and decoder 54, the same neighboring samples are available for a current block 18 both at encoder 14 and decoder 54. Accordingly, in order to avoid unnecessary repetition, the description of the mode of operation of encoder 14 shall also apply to decoder 54 as far the subdivision of picture 10 into blocks is concerned, for instance, as far as prediction is concerned and as far as the coding of the prediction residual is concerned. Differences lie in the fact that encoder 14 chooses, by optimization, some coding options or coding parameters and sig-nals within, or inserts into, datastream 12 the coding parameters which are then derived from the datastream 12 by decoder 54 so as to redo the prediction, subdivision and so forth.

Fig. 4 shows a possible implementation of the decoder 54 of Fig. 3, namely one fitting to the implementation of encoder 14 of Fig. 1 as shown in Fig. 2. As many elements of the encoder 54 of Fig. 4 are the same as those occurring in the corresponding encoder of Fig.

2, the same reference signs, provided with an apostrophe, are used in Fig. 4 in order to indicate these elements. In particular, adder 42', optional in-loop filter 46' and predictor 44' are connected into a prediction loop in the same manner that they are in encoder of Fig. 2. The reconstructed, i.e. dequantized and retransformed prediction residual signal applied to added 42' is derived by a sequence of entropy decoder 56 which inverses the entropy encoding of entropy encoder 28b, followed by the residual signal reconstruction stage 36' which is composed of dequantizer 38' and inverse transformer 40' just as it is the case on encoding side. The decoder's output is the reconstruction of picture 10. The reconstruction of picture 10 may be available directly at the output of adder 42' or, alterna- tively, at the output of in-loop filter 46'. Some post-filter may be arranged at the decoder's output in order to subject the reconstruction of picture 10 to some post-filtering in order to improve the picture quality, but this option is not depicted in Fig. 4.

Again, with respect to Fig. 4 the description brought forward above with respect to Fig. 2 shall be valid for Fig. 4 as well with the exception that merely the encoder performs the optimization tasks and the associated decisions with respect to coding options. However, all the description with respect to block-subdivisioning, prediction, dequantization and re-transforming is also valid for the decoder 54 of Fig. 4.

Before proceeding with the description of possible examples of the present application, some notes shall be made with respect to the above examples. Although not explicitly mentioned above, it is clear that block 18 may have any shape. It may be, for instance, of rectangular or quadratic shape. Moreover, although the above description of the mode of operation of encoder 14 and decoder 54 often mentioned a “current block” 18 it is clear that encoder 14 and decoder 54 act accordingly for each block for which an intra-prediction mode is to be selected. As described above, there may be other blocks as well, but the following description focuses on those blocks 18 into which picture 10 is subdivid-ed, for which an intra-prediction mode is to be selected.

In order to summarize the situation for a certain block 18 for which an intra-prediction mode is to be selected, reference is made to Fig. 5. Fig. 5 shows a current block 18, i.e. a block currently to be encoded or decoded. Fig. 5 shows a set 60 of neighboring samples 62, i.e. samples 62 with spatially neighbor block 18. The samples 64 within block 18 are to be predicted. The prediction signal to be derived is, thus, a prediction for each sample 64 within block 18. As already discussed above, a plurality 66 of prediction modes are avail-able for each block 18 and if block 18 is to be intra-predicted, this plurality 66 of modes merely comprises inter-prediction modes. A selection 68 is performed at encoder and de-coder side in order to determine one of the intra-prediction modes out of the plurality 66 to be used to predict (71) the prediction signal for block 18 on the basis of the neighboring sample set 60. The examples described further below differ with respect to the available intra-prediction modes 66 and the mode of operation with respect to selection 68 such as, for instance, whether side information is set in the datastream 12 with respect to selection 68 with respect to block 18 or not. The description of these examples, however, starts with a concrete description providing mathematical details. According to this initial example, the selection for a certain block 18 to be intra-predicted is associated with corresponding side information signalization 70 and the datastream and the plurality 66 of intra-prediction modes comprises a set 72 of neural network-based intra-prediction modes as well as a set 74 of further intra-prediction mode of heuristic design. One of the intra-prediction modes of set 74 may, for instance, be a DC prediction mode according to which some mean value is determined on the basis of the neighboring sample set 60 and this mean value is assigned to all samples 64 within block 18. Additionally or alternatively, set 74 may comprise inter-prediction modes which may be called angular inter-prediction modes according to which sample values of the neighboring sample set 60 are copied into block 18 along a certain intra-prediction direction with this intra-prediction direction differing among such angular intra-prediction modes. Fig. 5 shows that the datastream 12 corn-prises, in addition to the optionally present side information 70 concerning the selection 68 out of the plurality 66 of intra-prediction modes, a portion 76 into which the prediction re-sidual encoded which coding may, as discussed above, optionally involve transform cod-ing with quantization in transform domain.

In particular, in order to ease the understanding of the following description of a specific example of the present application, Fig. 6 shows the general mode of operation for an intra-prediction block at encoder and decoder. Fig. 6 shows block 18 along with the neighboring samples set 60 on the basis of which the intra-prediction is performed. It should be noted that this set 60 may vary among the intra-prediction modes of the plurality 66 of intra-prediction modes in terms of cardinality, i.e. the number of samples of set 60 actually used according to the respective intra-prediction mode for determining the predic-tion signal for block 18. This is, however, for ease of understanding, not depicted in Fig. 6. Fig. 6 shows that encoder and decoder have one neural network 800 to 80KB-1 for each of the neural network-based intra-prediction modes of set 72. Set 60 is applied to the respec-tive neural network so as to derive the corresponding intra-prediction mode among set 72. Besides this, Fig. 6 rather representatively shows one block 82 as providing on the basis of the input, namely the set 60 of neighboring samples, the one or more prediction signals of the one or more intra-prediction modes of set 74, e.g. the DC mode prediction signal and/or angular intra-prediction mode prediction signal. The following description will show as to how the parameters for the neural networks 80i with i=0... KB-1 may advantageously be determined. The specific example set out hereinafter, also provides encoder and de-coder with another neural network 84 which is dedicated to provide a probability value for each neural network-based intra-prediction mode within set 72 on the basis of a set 86 of neighboring samples which may or may not coincide with set 60. The probability values thus provided when the neural network 84 assists in rendering the side information 70 for the mode selection more effective. For instance, in the example described below, it is as-sumed that a variable length code is used to point to one of the intra-prediction modes and at least as far as set 72 is concerned, the probability values provided by the neural network 84 enable to use the variable length code within the side information 70 as an index into an ordered list of intra-prediction modes ordered according to the probability values output by neural network 84 for the neural network-based intra-prediction modes within set 72, thereby optimizing or reducing the code rate for the side information 70. To this, as depicted in Fig. 6, the mode selection 68 is effectively performed depending on both, the probability values provided by the further neural network 84 as well as the side information 70 within datastream 12.

1. Algorithm to train the parameters of Neural Networks that perform an intra pre-diction

Let B ⊂ ℤ2 be a block of a video frame, i.e. block 18. Assume that B has M pixels. For a fixed color component, let im be the content of a video signal on B. We regard im as an element of ℝM. Assume that there exists a neighbourhood Brec ⊂ ℤ2 of B that has L pixels and on which an already reconstructed image rec ∈ ℝL is available, i.e. sample sets 60 and 86 although they may alternatively differ. By an intra-prediction-function, we mean a function F: ℝL → ℝM. We regard F(rec) as a predictor for im.

What is described next is an algorithm to design, via a data-driven optimization approach, intra-prediction-functions for several blocks B that may occur in a typical hybrid video cod-ing standard, namely set 72. In order to achieve that goal, we took the following main de-sign features into account:

1. In the optimization algorithms that we conduct, we want to use a good ap-proximation of the cost function that in particular involves the number of bits one can ex-pect to spent to signal the prediction residual.

2. We want to train several intra predictions jointly in order to be able to handle different signal characteristics.

3. When training intra predictions, one has to take into account the number of bits needed to signal which intra mode is to be used.

4. We want to keep a set of already defined intra predictions, for example the HEVC intra predictions, and train our predictions as complementary predictions.

5. A typical hybrid video coding standard usually supports several blocks shapes into which the given block B can be partitioned.

In the next four sections, a possibility is to describe how one may deal with each of these requirements. More precisely, in section 1.1, we shall describe how to deal with the first item. In section 1.2, it is described how to handle items 2 to 3. In section 1.4, it is described how to take item 4 into account. Finally, in section 1.5, it is described how to deal with the last item.

1.1 Algorithm to train a loss function that approximates the Rate function of a vid-eo codec

A data driven approach to determine unknown parameters that are used in a video codec is usually set up as an optimization algorithm that tries to minimize a predefined loss func-tion on a given set of training examples. Typically, for a numerical optimization algorithm to work in practice, the latter loss function should satisfy some smoothness requirements.

On the other hand, a video encoder like HEVC performs best when it makes its decisions my minimizing the Rate-Distortion costs D + λ · R, Here, D is the reconstruction error of the decoded video signal and R is the rate, i.e. the number of bits needed to code the vid-eo signal. Moreover, λ ∈ ℝ is a Lagrangian Parameter that depends on the chosen Quan-tization Parameter.

The true function D + λ · R is typically very complex and is not given by a closed expres-sion one can feed a data driven optimization algorithm with. Thus, we approximate either the whole function D + λ · R or at least the rate function R by a piecewise smooth function.

More precisely, as before let B be a given block 1/ of a video frame 10 and let im be the corresponding video signal on B in a fixed color component. Assume that B has M pixels. Then for a prediction candidate pred ∈ ℝM, we consider the prediction residue res: =

(im ‒ pred) ∈ ℝM . For a given Quantization Parameter and a given transform, let R(res ) be the rate that a true video encoder needs to signal the quantized transform of res.

Moreover, let D (res) be the reconstruction error that arises by dequantization and inverse transform of res. Then we want to determine functions H,
which are piecewise smooth such that H(res) serves as a good approximation of D(res) + λ · R(res) and such that
serves as a good approximation of R(res).

We fix some and fix predefined "architectures", i.e. piecewise smooth functions

H: ℝM × ℝN → ℝ, R: ℝM × ℝN → ℝ

and then seek Φ1, Φ2 ∈ ℝN, such that we model our functions H and
as

H(res) = H (res, Φ1),

In order to determine the weights Φ1 and Φ2, on a typical encoder that uses the given

hybrid video coding standard we collected a huge set of training examples of prediction

residues resj , and the corresponding Rate-Distortion values (D + λR)(resj) respec-

tively only the rate values R(resj ) for some finite large index set
Then we try to find

and Φ2 such that they minimize or at least make small the expressions

For that task, we usually use a (stochastic) gradient descent approach.

1.2 Training of predictions for a fixed block shape

In this section we describe the algorithm that we set up to design KB intra-predictions for a given block B 18, the ones of st 72, and area Brec 60 for already reconstructed samples.

We assume that we are given a predefined "architecture" of our predictions. By this we

mean that for some fixed T ∈ M we are given a function

FB: ℝL × ℝT → ℝM (1) and that we want to determine "weights" such that our intra predic-

tions are given as

where for rec ∈ ℝL we put

The following section provides details in this regard. The functions in (2) define the neural network 800- 80KB -1 in Fig. 6.

Next, we model the signalization cost for the intra modes that we try to design by using a second parameter-dependent function

Again, for ΨB ∈ ℝL, we define

by

Again, an example is given in section 1.3 with the function of (4) representing neural net-work 84 of Fig. 6.

We assume that we are given a function

This function, for instance, defines a VLC code length distribution used for side infor-mation 70. i.e. the code lengths assocaited by side information 70 with cad ponite more of set 72.

Then we define

by

For the time being, the k-th component

of shall model the number of bits needed to signal the k-th intra mode that we

train.

If is the function defined in section 2.1, for given reconstructed im-

age rec ∈ ℝL and original image im ∈ ℝM, we let

denote the smallest k ∈ {1, ...,KB} with the property that

for all l ∈ {1, ...,KB}.

Since M models the true number of bits for the singalization of an intra mode, its gradient is either zero or undefined. Thus, M allone does not suffice to optimize the weights ΨB via a gradient-descent based algorithm. Thus, we also invoke the cross entropy of an intra

mode by transforming the function into a probability distribution using the softmax-

function. We recall the definition of the latter function. For x ∈ ℝT let xi denote the i-th component of x. Then the softmax function
is defined as

For gradient updates, we will try to minimize the sum of the rate of the residue and the

cross entropy of the mode kopt with respect to the latter probability distribution. Thus we

define our loss function LossB for the block B as

where

Given the loss function in (5), we determine the weights by a data driven opti-

mization.Thus, if for a finite, large index set
we are given a set of training examples

of images imi on B and corresponding reconstructed images reci on Brec, we apply an optimization algorithm, for example based on the (stochastic) gradient descent method, to find weights
that minimize the expression

(6)

1.3 Specification of the Functions

In this section, we define the form of the functions
more precisely. Again, recall that some define neural networks 80 and 84. Each of these functions consists of a se-quence of compositions of functions which are either: 1) An affine transofrmation Aff or

2) A non-linear activation function Act.

By an affine transformation Aff·. ℝm → ℝn, we mean a map that is of the form

Aff(x) = L(x) + b,

where L: ℝm → ℝn is a linear transformation, i.e. satisfies

L(λ · x1 + x2) =λ · L(x1) + L(X2)

for all λ ∈ ℝ, x1, x2 ∈ ℝm, and where b ∈ ℝn. Each linear map L: ℝm → ℝn is completely determined by a matrix in ℝn×m, i.e. corresponds uniquely to a vector ΘL ∈ ℝm·n. Each

affine function Aff: ℝm → ℝn is thus completely determined by m · n + n weights, i.e. by a vector Θ ∈ ℝm·n+n. For each Θ ∈ ℝm·n+n we shall write AffΘ for the unique affine trans-formation that corresponds to Θ in the aforementioned way.

By a non-linear activation function Act: ℝn → ℝn, we mean a function of the form

(Act(x))i = ρ(xi).

Here, (Act(x))i denotes the i-th component of Act(x) and xt denotes the i-th component of x. Finally, ρ: ℝ → ℝ my be of the form

or of the form

although these examples shall not be interpreted as limiting examples of the present ap-plication to these explicit examples. Other formulae may be used as well such as ρ(z) =

log(1 + ez) or any other non-linear function. ρ: ℝ → ℝ may alternatively be a piecewise

smooth function, for example.

Our function
now looks as follows. We assume that for a fixed we are

given
with m1 = L, nk = M, such that

T = (m1 · n1 + n1) + (m2 · n2 + n2) + ··· + (mk · nk + nk).

Here,
are as in (1). Then, for

with Θ = (Θ1, ... , Θk) ∈ ℝT, we define

would, thus, describe a neural network 80i parametrized using paramters Θ. It would

be a sequence of linear functions and non-linear functions ρ, which, in the present

example, are applied alternatingly in the sequence, wherein the parameters Θ comprise

the linear function weights in In the sequence of linear functions and non-

linear functions ρ, the pairs of a linear function followed by non-linear function p

would represent a neuron layer, for example, such as the j-th layer, with the number of predecessor nodes preceding this neuron layer j in feed-forward direction of the neural network being determined by dimension m of the number of columns of and

the number of neurons of the neuron layer j itself being determined by dimension n of the number of its rows. Each row of incorpartes the weights controlling as to

how strong a signal strength respectively activation of each of the m predecessor neurons is forwarded to the respective neuron of the neuron layer which corresponds to the re- spective row. ρ contrails for each neuron of neuron layer j the non-linear mapping of its linear combination of forwarded predecessor neuron activations onto its own activation. In the above example, there are k such neuron layers. The number of neurons per layer may vary. The number of nuron layers k may vary among the various neural networks 80j, i.e. for different j. Note, that the non-linear function might vary per neurion layer or even per neuron or at some other units.

Similarly, our function
looks as follows. We assume that for a fixed
we are given
with m1' = L, nk' = KB' such that

T = (m'1 · n'1 + n'1) + (m'2 · n'2 + n'2) + ··· + (m'k' · n'k' + n' k').

Here,
are as in (3). Then, for

Ψ =

( Ψ1, ... , Ψk') ∈ ℝT, we define

would, thus, describe a neural network 84 parametrized using paramters Ψ. It would be a sequence of linear functions and non-linear functions ρ, just as it has been

described above with respect to the neuron layers concerning the prediction signal com- putation. The number of neuron layers k' of neural network 84 may differ from one or more of the number of neuron layers k of neural networks 80i.

1.4 Training that takes existing predictions into account

We extended the algorithm of the previous section so that we can train predictions that complement already existing intra predictions.

Namely, let

be a set of fixed intra prediction functions that are already available. For example,
can consist of the DC- or Planar-prediction of HEVC and angular predictions defined ac- cording to HEVC; all those predictions may also include a preliminary smoothing of the reconstructed samples. Moreover, we assume that we are given a function

such that Lfix(im, rec, k) models the loss of the k-th intra prediciont function applied

to rec given the original image im.

Then we extend the loss function from (5) to the loss function

Keeping the notations from the end of the previous section, we determine weights

by minimizing

on a large set of training examples.

For that purpose, we typically firstly find the weights by optimizing (6) and then initialize

with those weights to find the weights that optimize (10).

1.5 Joint training of predictions for several block shapes

In this section we described how, in the training of our predictions, one may take into ac- count that in a typical video coding standard it is usually possible to split a block into

smaller subblocks in various ways and to perform an infra prediction on the smaller sub- blocks.

Namely, assume that for some
we are given a set

BL = {B1, ... , BS}

of admissible blocks Bi ⊂ ℤ2 together with a set of areas

such that each is a neighborhood of Bi. Typically, is a union of two rectan-

gles left and above Bi.

We assume that there exists a block Bmax ∈ BL such that Bi ⊆ Bmax for each i ∈ {1, ... , S}.

Let P(BL) be the power set of BL. Then for B ∈ BL we assume that a set

BL(B) ⊆ P(BL)

is given such that for each the block B can be written as a dis-

joint union

For a given color component, let im be an image on Bmax, which, by restriction, we regard as an image on Bi for each Bi ∈ BL. Moreover, assume that there exists a recon-

structed image rec on
which, by restriction, we regard as an image on
for each

Keeping the notations of section 1.2, for each B ∈ BL we seek as the set
of weights for KB infra prediction-functions and we seek ΨB ∈ ℝT as weights for the

mode prediction function GB. We determine these weights for all B ∈ BL jointly as follows. For B ∈ BL and given sets of weights
B' ∈ BL, B' ⊆ B, 1 ≤ k ≤ KB', we put

Moreover, for B' ⊂ B we define ΘB |B' ⊂ ΘB as

As in section 1.4, we assume that for each B ∈ BL a possibly empty set of intra pre-

diction functions is available. We let

Then we define a loss-function as follows. We have an ordering ≤ on the set

BL via the inclusion of sets. Let

BLmin ⊆ BL

be the set of all minimal elements in BL. For B ∈ BLmin we put

where the latter function is as in (9).

Next, let B ∈ BL and assume that LossB,total is already defined for all B' ∈ BL with B' ⊆ B.

Then we define

Finally, given a fixed set of training examples

of images imi on Bmax, we determine by minimizing or at least making small

the expression

We typically initialize the weights ΨB by firstly minimizing (9) for each B ∈ BL individ-

ually.

2 Integration of trained Neural Networks into a video codec

We consider a hybrid video coding standard in which for a given color component the con-tent of a video signal on a given block B ⊂ ℤ2 is to be generated by a decoder. Let M be the number of pixels of B. Moreover, let Brec ⊂ ℤ2 be a fixed neighbourhood of B such that the decoder has at its disposal a reconstructed image rec on Brec. Let L be the number of pixels of Brec. Then we regard rec as an element of ℝL. We assume that the codec oper-ates by predictive coding on the current block B 10. Then we claim copyright for the fol-lowing steps that a decoder can perform in order to generate a prediction signal pred on B, which we regard as an element of ℝM:

1. The decoder has at its disposal fixed numbers KB,
functions

FB: ℝL × ℝT → ℝM, namely 801 ... 80(CB-1) and
namely 84, as well as weights and a weight Ψ ∈ ℝT, where the latter weights are de-

termined in advance by a training algorithm that was described in the previous section.

2. The decoder reconstructs from the bitstream a flag that is part of side infor-mation 70 and indicates whether exactly one of the following options is true: [label=)]

(i) One of the predictions is to be used, i.e. a mode out of set

72

(ii) None of the predictions is to be used, i.e. one out of 74, for instance

Here, the functions are as in (2) ,

3. If Option Two in step 2 is true, the decoder proceeds for the given block 10 as in the underlying hybrid video coding standard.

4. If Option One in step 2 is true, the decoder applies the function
i.e. 84, defined according to (4), to the reconstructed image rec. Let be
defined as
Then the standard is changed in a way such that the decoder defines a number m ∈ {1, ... ,KB} by exactly one of the following tow options:

(i) The decoder defines a probability distribution on the set {1, ...,KB}
by

and uses the latter probability distribution to parse an index k ∈ {1, ... ,KB} that is

also part of side information 70 via the entropy coding engine used in the underlying standard from the datastream 12 and defines m: = k.

(ii) The decoder defines a permutation

σ: {1, ... ,KB} → {1, ... ,KB}

inductively by putting
where is the minimal number with

for all k ∈ {1, ... KB} and by Putting
where
is the minimal number such that one has for all k ∈ {1, ... ,KB}\

(σ(1), ..., σ(l)}.

Then the decoder reconstructs from the bitstream 12 a unique index i ∈ {1, ... KB} that Is also part of datastream 12 and puts m: = σ(i).

In the code design to parse the latter index i, it is required that the number of bits needed to signal an index i1 ∈ {1, ... ,KB} is less or equal than the number of bits to signal an index i2 ∈ {1, ... ,KB} if σ(i1) ≤ σ(i2) and if all involved underlying probabilities used by the entropy coding engine are set to equal probability.

5. If Option One in step 2 is true and if the decoder has determined the index m according to the previous step 4, the decoder generates 71 the prediction signal pred ∈ RM as i.e. using the selected neural network 80m. Then the decoder

proceeds as in the underlying hybrid video coding standard using pred as prediction sig- nal.

The integration of intra prediction functions whose design is based on a data driven learn-ing approach into an existing hybrid video codec. The description had two main parts. In the first part, we described a concrete algorithm for an offline training of intra prediction functions. In the second part, we described how a video decoder may use the latter pre-diction functions in order to generate a prediction signal for a given block.

Thus, what has been described above in sections 1.1 to 2, is, inter alia, an apparatus for block-wise decoding a picture 10 from a datastream 12. The apparatus 54 supports a plu-rality of intra-prediction modes comprising, at least, a set 72 of intra-prediction modes ac-cording to which the intra-prediction signal for a current block 18 of the picture 10 is de-termined by applying a first set 60 of neighboring samples of the current block 18 onto a neural network 80i. The apparatus 54 is configured to select (68) for the current block 18 one intra-prediction mode out of the plurality 66 of intra-prediction modes and predict (71) the current block 18 using the one intra-prediction mode, namely using the corresponding neural network 80m having been selected. Although the decoder presented in section 2, had intra-prediction modes 74 within the plurality 66 of intra-prediction modes supported in addition to the neural network-based ones in set 72, this has been merely an example and needs not to be the case. Further, the above description in sections 1 and 2 may be var-ied in that decoder 54 does not use, and does not comprise, the further neural network 84. With respect to the optimization described above, this means that the second adder in the inner quality presented in section 1.2 for finding-out would not have to be a concate-

nation of a function MB applied onto any probability value neural network function GB. The optimization algorithm of what, rather, determines suitable parameters for the neural net-works 80j in a manner so that the frequency of selection would appropriately follow a code rate indication of MB. For instance, the decoder 54 could decode from datastream 12 an index for block 18 using a variable length code, the code length of which are indicated in MB, and the decoder 54 would perform the selection 68 based on this index. The index would be part of the side information 70.

A further alternative to the description brought forward above in section 2 is that the de-coder 54 may alternatively derive a ranking among the set 72 of neural network-based intra-prediction modes depending on a first portion of the datastream which relates to a neighborhood of the current block 18 in order to obtain an ordered list of intra-prediction modes with selecting the intra-prediction mode finally to be used out of the ordered list of intra-prediction modes depending on a second portion of the datastream other than the first portion. The “first portion” may, for instance, relate to a coding parameter or prediction parameter related to one or more block neighboring current block 18. The “second por-tion” may then be an index, for instance, pointing into, or being an index of, the neural network-based intra-prediction mode set 72. When construed in alignment with above-outlined section 2, the decoder 54 comprises the further neural network 84 which deter-mines, for each intra-prediction mode of the set 72 of intra-prediction modes, a probability value by applying set 86 of neighboring samples thereonto and ordering these probability values in order to determine a rank for each intra-prediction mode of set 72, thereby ob-taining an ordered list of intra-prediction modes. An index in the datastream 12 as part of side information 70 is then used as an index into the ordered list. Here, this index may be coded using variable length code for which MB indicates the code length. And as ex-plained above in section 2, in item 4i, according to a further alternative example, decoder 54 may use the just-mentioned probability values determined by the further neural net-work 84 for each neural network-based intra-prediction mode of set 72 so as to efficiently perform entropy coding of the index into set 72. In particular, the symbol alphabet of this index which is part of the side information 70 and used as an index into set 72, would comprise a symbol or value for each of the modes within set 72, and the probability values provided by neural network 84 would, in case of neural network 84 design according to the above description, provide probability values which would lead to efficient entropy cod-ing in that these probability values closely represent the actual symbol statistics. For this entropy coding, arithmetic coding could be used, for instance, or probability interval parti-tioning entropy (PIPE) coding.

Favorably, no additional information is necessary for any of the intra-prediction modes of set 72. Each neural network 80i, once advantageously parametrized for encoder and de-coder in accordance with, for example, the above description in sections 1 and 2, derives the prediction signal for the current block 18 without any additional guidance in the datastream. As already denoted above, the existence of other intra-prediction modes be-sides the neural network-based ones in set 72 is optional. They have been indicated above by set 74. In this regard, it should be noted that one possible way of selecting set 60, i.e. the set of neighboring samples forming the input for the intra-prediction 71, may be such that this set 60 is the same for the intra-prediction modes of set 74, i.e. the heuristic ones, with set 60 for the neural network-based intra-prediction modes being larger in terms of the number of neighboring samples included in set 60 and influencing the intra-prediction 71. In other words, the cardinality of set 60 may be larger for neural network-based intra-prediction modes 72 compared to the other modes of set 74. For instance, set 60 of any intra-prediction mode of set 74 may merely comprise neighboring samples along

a one-dimensional line extending alongside to sides of block 18 such as the left hand one and the upper one. Set 60 of the neural network-based intra-prediction modes may cover an L-shaped portion extending alongside the just-mentioned sides of block 18 but being wider than just one-sample wide as set 60 for the intra-prediction modes of set 74. The L shaped portion may additionally extend beyond the just mentioned sides of block 18. In this manner, neural network-based intra-prediction modes may result into a better intra-prediction with a correspondingly lower prediction residual.

As described above in section 2, the side information 70 conveyed in the datastream 12 to an intra-predicted block 18 may comprise a fleck which generally indicates whether the selected intra-prediction mode for block 18 is member of set 72 or member of set 74. This fleck is, however, merely optional with side information 70 indicating, for instance, an in-dex into a whole plurality 66 of intra-prediction modes including both sets 72 and 74.

The just-discussed alternatives are, in the following, briefly discussed with respect to the Figs. 7a to 7d. The Figs, define both, decoder and encoder concurrently, namely in terms of their functionality with respect to an intra-predicted block 18. The differences between the encoder mode of operation and the decoder mode of operation with respect to an in-tra-coded block 18 is, on the one hand, the fact that the encoder performs all or at least some of the intra-prediction modes 66 available so as to determine at 90 a best one in terms of, for instance, some cost function minimizing sense, and that the encoder forms data stream 12, i.e., codes date there into, while the decoder derives the data therefrom by decoding and reading, respectively. Fig. 7a shows the mode of operation for the above-outlined alternative according to which a flag 70a within the side information 70 for block 18 indicates whether the intra-prediction mode determined to be the best mode for block 18 by the encoder in step 90, is within set 72, i.e., is neural network based intra-prediction mode, or within set 74, i.e., one of the non-neural network based intra-prediction modes. The encoder inserts flag 70a into data stream 12 accordingly, while the decoder retrieves it therefrom. Fig. 7a assumes that the determined intra-prediction mode 92 is within set 72. The separate neural network 84 then determines a probability value for each neural network based intra-prediction mode of set 72 and using these probability values set 72 or, to be more precise, the neural network based intra-prediction modes therein are ordered according to their probability values such as in descending order of their probability values, thereby resulting into an ordered list 94 of intra-prediction modes. An index 70b being part of the side information 70 is then coded by the encoder into data stream 12 and decoded therefrom by the decoder. The decoder, accordingly, is able to

determine which set of sets 72 and 74. The intra-prediction mode to be used for block 18 is located in, and to perform the ordering 96 of set 72 in case of the intra-prediction mode to be used being located in set 72. It might be that an index is also transmitted in data stream 12 in case of the determined intra-prediction mode being located in set 74. Thus, the decoder is able to generate the prediction signal for block 18 using the determined intra-prediction mode by controlling the selection 68 accordingly.

Fig. 7b shows an alternative according to which the flag 70a is not present in data stream 12. Instead, the ordered list 94 would not only comprise the intra-prediction modes of set 72, but also intra-prediction modes of set 74. The index within side information 70 would be an index into this greater ordered list and indicate the determined intra-prediction mode, i.e., the one determined be optimization 90. In case of neural network 84 providing a probability value for the neural network based intra-prediction mode within 72 only, the ranking between intra-prediction modes of set 72 relative to the intra-prediction modes of set 74 may be determined by other means such as inevitably arranging the neural network based intra-prediction modes of set 72 to precede the modes of set 74 in the order list 94 or to arrange them alternatingly relative to each other. That is, the decoder is able to de-rive the index from data stream 12, use the index 70 as in index into the order list 94 with deriving the order list 94 from the plurality of intra-prediction modes 66 using the probabil-ity values output by neural network 84. Fig. 7c shows a further variant. Fig. 7c show a case of not using flag 70a, but the flag could be used alternatively. The issue which Fig.

7c is directed pertains to the possibility that neither encoder nor decoder uses neural net-work 84. Rather, the ordering 96 is derived by other means such as coding parameters conveyed within data stream 12 with respect to one or more neighboring blocks 18, i.e., portions 98 of a data stream 12 which pertains to such one or more neighboring blocks.

Fig. 7d shows a further variant of Fig. 7a, namely the one according to which the index 70b is coded using entropy coding and decoded from data stream 12 using entropy de-coding, commonly denoted using reference sign 100. The sample statistics or the proba-bility distribution used for the entropy coding 100 is controlled by the probability values output by neural network 84 as explained above, this renders the entropy coding of index 70b very efficient.

For all examples 7a to 7d it is true that set 74 modes may not be present. Accordingly, the respective module 82 may be missing and flag 70a would be unnecessary anyway.

Further, although not shown in any Fig., it is clear that the mode selection 68 at the en- coder and decoder could be synchronized to each other even without any explicit signal- ing 70, i.e., without spending any side information. Rather, the selection could be derived from other means such as by taking inevitably the first one of the ordered list 94, or by deriving the index into the order list 94 on the basis of coding parameters relating to one or more neighboring blocks. Fig. 8 shows an apparatus for designing the set of intra- prediction modes of set 72 to be used for the block-based picture coding. The apparatus 108 comprises a parameterizable network 109 which inherits or comprises parameteriza- ble versions of neural networks 800 to 80KB-1 as well as neural network 84. Here, in Fig. 8, depicted as individual units, i.e., neural network 840 for providing the probability value for neural network based intra-prediction mode 0 to neural network 84KB-1 for providing the probability value associated with the neural network based intra-prediction mode KB-1. The parameters 111 for parametrizing neural networks 84 and the parameters 113 for para- metrizing neural networks 800 to 80KB-1 are input or applied to respective parameter inputs of these neural networks by an updater 110. Apparatus 108 has access to a reservoir or a plurality of picture test blocks 114 along with corresponding neighboring samples sets 116. Pairs of these blocks 114 and their associated neighboring sample sets 116 are se- quentially used by apparatus 108. In particular, a current picture test block 114 is applied to parameterizable neural network 109 so that neural networks 80 provide a prediction signal 118 for each neural network based intra-prediction mode of set 72, and each neural network 80 provides a probability value for each of these modes. To this end, these neural networks use their current parameters 111 and 113.

In the above description rec has been used to denote the picture test block 114, and

is the prediction residual 118 for mode B and the probability value is is

the probability value 120. For each mode 0... Kb-1, there is a cost estimator 122 comprised by apparatus 108 which computes a cost estimate for the respective mode on the basis of the prediction signal 118 obtained for the respective mode. In the above example, cost estimators 122 computed the cost estimates as indicated on the left and right hand sides of the inequality in section 1.2. That is, here, the cost estimators 122 also used, for each mode, the corresponding probability value 120. This needs not, however, to be case as already discussed above. The cost estimate, however, is in any case a sum of two add- ins, one of which is an estimate of the coding cost for the prediction residual indicated as the term with
in the above inequality, and another add-in estimating the coding costs for indicating the mode. In order to compute the estimate for the coding cost related to the prediction residual, the cost estimators 122 also obtain the original content of the current picture test block 114. The neural networks 80 and 84 had at their inputs applied thereto the corresponding neighboring sample sets 116. The cost estimate 124 as output by cost estimators 122 is received by a minimum cost selector 126 which determines the mode minimizing or having minimum cost estimate associated therewith. In the above mathe-matical notation, this has been The updater receives this optimum mode and uses a

coding cost function having a first add in forming residual rate estimate depending on the prediction signal 118 obtained for the intra-prediction mode of lowest coding estimate, and a second add-in forming a mode signaling side information rate estimate depending on the prediction signal and the probability value obtained for the intra-prediction mode of lowest coding cost estimate as indicated by selector 126. As indicated above, this may be done using a gradient distant. The coding cost function is, thus, differentiable and in the above mathematical representation an example of this function was given in equation 5. Here, the second add-in relating to the mode signaling side information rate estimate computed the cross entropy for the intra-prediction mode of lowest coding cost estimate.

Thus, the updater 110 seeks to update parameters 111 and 113 so as to reduce the cod-ing cost function and then these updated parameters 111 and 113 are used by the para-metrizable neural network 109 so as to process the next picture test block of the plurality 112. As discussed above with respect to section 1.5, there may be a mechanism control-ling that primarily those pairs of picture test blocks 114 and their associated neighboring sample sets 116 are applied for the recursive update process for which the intra-prediction is, in rate distortion sense, preferably done without any block sub-division, thereby avoid-ing that the parameters 111 and 113 are optim1zed too much on the basis of picture test blocks for which, anyway, a coding in units of sub-blocks thereof is more cost effective.

So far, the above-discussed examples primarily concern cases where encoder and de-coder had within their supported intra-prediction modes 66 a set of neural network-based intra-prediction modes. In accordance with the examples discussed with respect to Figs.

9a and 9b this needs not to be the case necessarily. Fig. 9a seeks to outline the mode of operation of an encoder and a decoder in accordance with an example wherein the de-scription thereof is provided in a manner focusing on the differences to the description brought forward above with respect to Fig. 7a. The plurality 66 of supported intra-prediction modes may or may not comprise neural network-based intra-prediction modes and may or may not comprise non-neural network-based intra-prediction modes. The modules 170 in Fig. 9a comprised by encoder and decoder, respectively, in order to pro-vide for each of the supported modes 66 the corresponding prediction signal are, accord- ingly, not necessarily neural networks. As already denoted above, such intra-prediction modes may be neural network-based or they may be heuristically motivated and compute the prediction signal based on a DC intra-prediction mode or an angular intra-prediction mode or any other. Accordingly, these modules 170 could be denoted as prediction signal computer. Encoder and decoder in accordance with the example of Fig. 9a, however, comprise a neural network 84. The neural network 84 computes, on the basis of the neighboring sample set 86, probability values for the supported intra-prediction modes 66 so that the plurality 66 of intra-prediction modes may be turned into the ordered list 94. The index 70 within datastream 12 for block 18 points into this ordered list 94. The neural network 84, thus, assists in lowering the side information rate to be spent for the intra-prediction mode signalization.

Fig. 9b shows an alternative to Fig. 9a in that instead of the ordering, entropy de/encoding 100 of the index 70 is used with controlling the probability or simple statistics thereof, i.e. controlling the entropy probability distribution for entropy de/encoding in en/decoder, ac-cording to the probability values determined for the neural network 84 for each mode of plurality 66.

Fig. 10 shows an apparatus for designing or parametrizing neural network 84. It is, thus, an apparatus 108 for designing a neural network for assisting in selecting among a set 66 of intra-prediction modes. Here, for each mode of set 66 there is a corresponding neural network block together forming neural network 84 and the parametrizable neural network 109 of apparatus 108 is merely parametrizabie with respect these blocks. For each mode, there is also the prediction signal computer 170 which needs, however, not to be para-metrizabie according to Fig. 10. Thus, apparatus 108 of Fig. 10 computes costs estimates for each mode on the basis of the prediction signal 118 as computed by the corresponding prediction signal computer 170 and, optionally, on the basis of the corresponding probabil-ity value as determined by the corresponding neural network block for this mode. On the basis of the resulting costs estimates 124, the minimum cost selector 126 selects the mode of the minimum cost estimate and the updater 110 updates the parameters 111 for the neural 84.

The following is noted with respect to the description of Figs. 7a to 7d and 9a and 9b. A common feature of the examples of Figs. 9a and 9b which is also used by some of the examples of Figs. 7a to 7d was the fact that the probability values of the neural network values in order to improve or reduce the overhead associated with the side information 70 for signaling the mode determined on the encoder side at the optimization process 90 to the decoder. As indicated above with respect to the examples of Figs. 7a to 7d, however, it should be clear that the examples of Figs. 9a and 9b may be varied to the extent that no side information 70 is spent in datastream 12 with respect to the mode selection at all. Rather, the probability values output by neural network 84 for each mode may be used to synchronize the mode selection between encoder and decoder inevitably. In that case, there would be no optimization decision 90 at the encoder side with respect to the mode selection. Rather, the mode to be used among set 66 would be determined on encoder and decoder side in the same manner. A similar statement is true with respect to corre-sponding examples of Figs. 7a to 7d when varied so as to not use any side information 70 in datastream 12. Back to the examples of Figs. 9a and 9b, it is interesting, however, that while the selection process 68 at the decoder side is dependent on the probability values output by the neural network in that the ordering or the probability distribution estimate dependency on the probability values varies the interpretation of the side information, as far as the encoder is concerned, the dependency on the probability values may not only affect the coding of the side information 70 into datastream 12 using, for instance, a re-spective variable length coding of an index into the ordered list or using entropy en/decoding with a probability distribution estimation depending on the neural network's probability values, but also the optimization step 90: here, the code rate for transmitting side information 70 may be taken into account and may, thus, influence the determination 90.

Example of Fig. 11-1

Fig. 11-1 shows a possible implementation of an encoder 14-1, namely one where the encoder is configured to use transform coding for encoding the prediction residual alt-hough this is nearly an example and the present application is not restricted to that sort of prediction residual coding. According to Fig. 11-1, encoder 14-1 comprises a spatial-domain subtractor 22 configured to subtract from the inbound signal, i.e. picture 10 or, on a block basis, current block 18, the corresponding prediction signal 24-1 so as to obtain the spatial-domain prediction residual signal 26 which is then encoded by a prediction residual encoder 28 into a datastream 12. The prediction residual encoder 28 comprises a lossy encoding stage 28a and a lossless encoding stage 28b. The lossy encoding stage 28a receives the prediction residual signal 26 and comprises a quantizer 30 which quan-tizes the samples of the prediction residual signal 26. The present example uses trans-form coding of the prediction residual signal 26 and accordingly, the lossy encoding stage 28a comprises a transform stage 32 connected between subtractor 22 and quantizer 30

so as to transform such a spectrally decomposed prediction residual 27 with a quantiza-tion of quantizer 30 taking place on the transformed coefficients where presenting the re-sidual signal 26. The transform may be a DCT, DST, FFT, Hadamard transform or the like. The transformed and transform-domain quantized prediction residual signal 34 is then subject to lossless coding by the lossless encoding stage 28b which is an entropy coder entropy coding quantized prediction residual signal 34 into datastream 12.

Encoder 14-1 further comprises a transform-domain prediction residual signal reconstruc-tion stage 36-1 connected to the transform-domain output of quantizer 30 so as to recon-struct from the transformed and quantized prediction residual signal 34 (in the transform domain) the prediction residual signal in a manner also available at the decoder, i.e. tak-ing the coding loss of quantizer 30 into account. To this end, the prediction residual re-construction stage 36-1 comprises a dequantizer 38-1 which performs the inverse of the quantization of quantizer 30 to obtain a dequantized version 39-1 of the prediction residual signal 34, followed by an inverse transformer 40-1 which performs the inverse transfor-mation relative to the transformation performed by transformer 32 such as the inverse of the spectral decomposition such as the inverse to any of the above-mentioned specific transformation examples. Downstream to the inverse transformer 40-1 , we have a spatial-domain output 60 which may comprise a template which will help to obtain the prediction signal 24-1. In particular, the predictor 44-1 may provide a transform-domain output 45-1 which, once inverse-transformed at the inverse transformer 51-1, will provide the predic-tion signal 24-1 in the spatial domain (the prediction signal 24-1 will be subtracted from the inbound signal 10, to obtain the prediction residual 26 in the time domain). There is also the possibility that, in inter-frame modes, an in-loop filter 46-1 filters completely re-constructed pictures 60 which, after having been filtered, form reference pictures 47-1 for predictor 44-1 with respect to inter-predicted block (accordingly, in these cases an adder 57-1 input from the elements 44-1 and 36-1 is necessary, but there is no necessity for the inverse transformer 51-1, as indicated by the dotted line 53-1, for providing the prediction signal 24-1 to the subtractor 22).

Differently from encoder 14 of Fig. 2, however, encoder 14-1 (at the prediction residual reconstruction stage 36-1) comprises a transform-domain adder 42-1, which is positioned between the dequantizer 38-1 and the inverse transformer 40-1. The transform-domain adder 42-1 provides the inverse transformer 40-1 with a sum 43-1 (in the transform do-main) of the dequantized version 39-1 of the prediction residual signal 34 (as provided by the dequantizer 38-1) with a transform-domain prediction signal 45-1 as provided by a

transform-predictor 44-1. The predictor 44-1 may obtain, as a feedback input, the output from the inverse transformer 40-1.

Hence, the prediction signal 24-1 in the spatial domain is obtained from a prediction signal 45-1 in the transform domain. Also the transform-domain predictor 44-1, which may oper-ate with neural networks according to the examples above, is input by signals in the spa-tial domain but outputs signals in the transform domain.

Example of Fig. 11-2

Fig. 11-2 shows a possible implementation of a decoder 54-2, namely one fitting to the implementation of encoder 14-1. As many elements of the encoder 54-2 are the same as those occurring in the corresponding encoder of Fig. 11-1, the same reference signs, pro-vided with a “-2”, are used in Fig. 11-2 in order to indicate these elements. In particular, adder 42-2, optional in-loop filter 46-2 and predictor 44-2 are connected into a prediction loop in the same manner that they are in encoder of Fig. 11-1. The reconstructed, i.e. dequantized and retransformed prediction residual signal 24-2 (e.g., 60), is derived by a sequence of entropy decoder 56 which inverses the entropy encoding of entropy encoder 28b, followed by the residual signal reconstruction stage 36-2 which is composed of dequantizer 38-2 and inverse transformer 40-2 just as it is the case on encoding side. The decoder’s output is the reconstruction of picture 10. Some post-filter 46-2 may be ar-ranged at the decoder’s output in order to subject the reconstruction of picture 10 to some post-filtering in order to improve the picture quality. Again, the description brought forward above with respect to Fig. 11-1 shall be valid for Fig. 11-2 as well with the exception that merely the encoder performs the optimization tasks and the associated decisions with respect to coding options. However, all the description with respect to block-subdivisioning, prediction, dequantization and retransforming is also valid for the decoder 54 of Fig. 11-2. The reconstructed signal 24-2 is provided to a predictor 44-2, which may operate with neural networks according to the examples of Figs. 5-10. The predictor 44-2 may provide a transform-domain prediction value 45-2.

Contrary to the example in Fig. 4, but similarly to that of Fig. 11-1 , the dequantizer 38-2 provides a dequantized version 39-2 of the prediction residual signal 34 (in the transform domain) which is not directly provided to the inverse transformer 40-2. Instead, the dequantized version 39-2 of the prediction residual signal 34 is input to an adder 42-2, so as to be composed with the transform-domain prediction value 45-2. Accordingly, there is obtained a transform-domain reconstructed signal 43-2, which, when subsequently in-

versely transformed by the inverse transformer 40-2, becomes the reconstructed signal 24-2 in the spatial domain, to be used for displaying the image 10.

Example of Fig. 12

Reference is now made to Fig. 12. Both, decoder and encoder concurrently, namely in terms of their functionality with respect to an intra-predicted block 18. The differences be-tween the encoder mode of operation and the decoder mode of operation with respect to an intra-coded block 18 is, on the one hand, the fact that the encoder performs all or at least some of the intra-prediction modes 66 available so as to determine at 90 a best one in terms of, for instance, some cost function minimizing sense, and that the encoder forms data stream 12, i.e., codes date there into, while the decoder derives the data therefrom by decoding and reading, respectively. Fig. 12 shows the mode of operation for the above-outlined alternative according to which a flag 70a within the side information 70 for block 18 indicates whether the intra-prediction mode determined to be the best mode for block 18 by the encoder in step 90, is within set 72, i.e., is neural network based intra-prediction mode, or within set 74, i.e., one of the non-neural network based intra-prediction modes. The encoder inserts flag 70a into data stream 12 accordingly, while the decoder retrieves it therefrom. Fig. 12 assumes that the determined intra-prediction mode 92 is within set 72. The separate neural network 84 then determines a probability value for each neural network based intra-prediction mode of set 72 and using these probability values set 72 or, to be more precise, the neural network based intra-prediction modes therein are ordered according to their probability values such as in descending order of their probability values, thereby resulting into an ordered list 94 of intra-prediction modes. An index 70b, being part of the side information 70, is then coded by the encoder into data stream 12 and decoded therefrom by the decoder. The decoder, accordingly, is able to determine which set of sets 72 and 74. The intra-prediction mode to be used for block 18 is located in, and to perform the ordering 96 of set 72 in case of the intra-prediction mode to be used being located in set 72. It might be that an index is also transmitted in data stream 12 in case of the determined intra-prediction mode being located in set 74. Thus, the decoder is able to generate the prediction signal for block 18 using the determined intra-prediction mode by controlling the selection 68 accordingly.

As may be seen from Fig. 12, the prediction residual signal 34 (in the transform domain) is encoded in the datastream 12. The dequantizer 38-1, 38-2 derives a transform-domain dequantized prediction residual signal 39-1, 39-2. From the predictor 44-1, 44-2, the trans-form-domain prediction signal 45-1, 45-2 is obtained. Then, the adder 42-1 sums the val- ues 39-1 and 45-1 (or the adder 42-2 sums the values 39-2 and 45-2) with each other, to obtain the transform-domain reconstructed signal 43-1 (or 43-2). Downstream to the in-verse transformer 40-1 , 40-2, a spatial-domain prediction signal 24-1, 24-2 (e.g., template 60) is obtained and may be used to reconstruct the block 18 (which may be, for example, displayed).

Al the variants of Figs 7b-7d may be used to embody the examples of Figs. 11-1 , 11-2, and 12.

Discussion

A method to generate an intra prediction signal via a Neural Network is defined and it is described how this method is to be included into a video- or still-image codec. In these examples, instead of predicting into the spatial domain, the predictors 44-1, 44-2 may predict into the transform domain for a predefined image transform that might be already available at the underlying codec, e.g. the Discrete Cosine Transform. Second, each of the intra prediction modes that is defined for images on blocks of a specific shape induces intra prediction modes for images on larger blocks.

Let B be a block of pixels with M rows and N columns on which an image im is present. Assume that there exists a neighbourhood Brec (template 60 or 86) of B (block 18) on which an already reconstructed image rec is available. Then in the examples of Figs. 5-10, new intra prediction modes defined by a Neural Network are introduced. Each of these intra prediction modes generates, using the reconstructed samples rec (24-1 , 24-2), a prediction signal pred (45-1, 45-2), which is again an image on Brec.

Let T be an image transform (e.g., prediction residual signal 34 as output by element 30) that is defined on images on Brec and let S be the inverse transform of T (e.g., at 43-1 or 43-2). Then the prediction signal pred (45-1 , 45-2) is to be regarded as a prediction for T(im). This means that at the reconstruction stage, after the computation of pred (45-1 ,45-2) the image S(pred) (24-1 , 24-2) has to be computed to obtain the actual prediction for the image im (10).

It has been noted that the transform T we work with has some energy compaction proper-ties on natural images. This is exploited in the following way. For each of our intra modes defined by a Neural Network, by a predefined rule the value of pred (45-1, 45-2) at specif-ic positions in the transform domain is set to zero, independent of the input rec (24-1 , 24- 2). This reduces the computational complexity to obtain the prediction signal pred (45-1, 45-2) in the transform domain.

(With ref to Figs. 5-10, assume that the transform T (at 32) and the inverse transform S (at 40) are used in the transform-residual coding of the underlying codec. To obtain the reconstruction signal (24, 24’) on B, a prediction residual res (34) is to be inverse trans-formed by the inverse transform S (at 40) to obtain S(res) and that S(res) is to be added to an underlying prediction signal (24) to obtain the final reconstruction signal (24).)

In contrast, Figs. 11 and 12 refer to the following procedure: in the case that the prediction signal pred (45-1, 45-2) is generated by the Neural Network intra prediction method as above, the final reconstruction signal (24-1 , 24-2) is obtained by the inverse transform (at 40-1, 40-2) of pred+res (where pred is 45-1 or 45-2; and res is 39-1 or 39-2, and their sum is 43-1 or 43-2, which is the transform-domain version of the final reconstruction signal 24-1, 24-2).

We finally remark that above modifications of the intra predictions performed by Neural Networks as above are optional and non-necessarily interrelated to each other. This means that for a given transform T (at 32) with inverse transform S (at 40-1 , 40-2) and for one of the intra prediction modes defined by a Neural Network as above, it might be ex-tracted either from the bitstream or from predefined settings whether the mode is to be regarded as predicting into the transform domain corresponding to T or not.

Figs. 13a and 13b

With reference to Figs. 13a and 13b, a strategy is shown which may be applied, for exam-ple, to spatial-domain-based methods (e.g., Figs. 11a and 11b) and/or to transform-domain-based methods (e.g., Figs. 1-4).

In some cases, there is, at disposal, a neural network adapted for blocks of a particular size (e.g., MxN, where M is the number of rows and N is the number of columns), while the real block 18 of the image to be reconstructed has a different size (e.g., M1xN1). It has been noted that it is possible to perform operations which permit to make use of the neural network adapted for a particular size (e.g., MxN), without necessity of using neural net-works trained ad hoc.

In particular, the apparatus 14 or 54 may permit block-wise decoding a picture (e.g., 10) from a data stream (e.g., 12). The apparatus 14, 54 natively supports at least one intra-prediction mode, according to which the intra-prediction signal for a block (e.g., 136, 172) of a predetermined size (e.g., MxN) of the picture is determined by applying a first tern-plate (e.g., 130, 170) of samples which neighbors the current block (e.g., 136, 176) onto a neural network (e.g., 80). The apparatus may be configured, for a current block (e.g., 18) differing from the predetermined size (e.g., M1xN1), to:

- resample (e.g., D, 134, 166) a second template (e.g., 60) of samples neighboring the current block (e.g., 18), so as to conform with the first template (e.g., 130, 170) so as to obtain a resampled template (e.g., 130, 170),

- apply the resampled template (e.g., 130, 170) of samples onto the neural network (e.g., 80) so as to obtain a preliminary intra-prediction signal (e.g., 138), and

- resample (e.g., U, V, 182) the preliminary intra-prediction signal (138) so as to conform with the current block (18, Bi) so as to obtain the intra-prediction signal for the current block.

Fig. 13a shows an example in the spatial domain. A spatial-domain block 18 (also indicat-ed also as B1) may be an M1xN1 block for which an image im1 is to be reconstructed (even if, at the moment, the image im1 is not available yet). It is noted that a template B1,rec (e.g., a set 60) has an already reconstructed image rec1, wherein rec1 is neighbouring to im1 (and B1,rec is neighboring to B1). The block 18 and the template 60 (“second template”) may form the element 132.

There arises the possibility that there is no neural network at disposal for reconstructing B1, by virtue of the dimensions of B1. However, in case neural networks are at disposal for a block with different dimensions (e.g., a “first template"), the following procedure may be implemented.

A transformation operation (here indicated as D or 134) may, for example, be applied to the element 130. It has been noted, however, that it is simply possible to apply the trans-formation D (130) to B1,rec alone, by virtue of B1 being still unknown. The transformation 130 may provide an element 136, which is formed of a transformed (resampled) template 130 and a block 138.

For example, the M1xN1 block B1 (18) (with unknown coefficients) may be theoretically transformed into an MxN block B (138) (with still unknown coefficients). As the coefficients of block B (138) are unknown, however, there is no practical necessity for actually per-forming the transformation.

Analogously, the transformation D (134) transforms the template B1,rec (60) into a different template Brec (130) with different dimensions. The template 130 may be L-shaped, with vertical thickness L (i.e., L columns in the vertical portion) and horizontal thickness K (i.e., K rows in the horizontal portion), with Brec =D(B1,rec). It may be understood that the tem-plate 130 may comprise:

- a KxN block over Brec (130);

- an MxL block at the left of Brec (130); and

- a KxL block at the left of the KxN block over Brec (130), and over the MxL block at the left of Brec (130).

In some cases, the transformation operation D (134) may be, where Mi>M and N1>N (and in particular where M is a multiple of M1 and N is a multiple of N1), a downsampling opera-tion. For example, in case of M1=2M and N1=2N, the transformation operation D may simply be based on hiding some bins in a chess-like fashion (e.g., by deleting diagonals from B1,rec 60, to obtain the values of Brec 130).

At this point, Brec (with Brec= D(rec1)) is a reconstructed image in MxN. At passage 138a, apparatus 14, 54 may now use (e.g., at the predictor 44, 44’) the necessary neural net-work (e.g., by operating as in Figs. 5-10) natively trained for an MxN block. By applying the passages discussed above (138a), the image im1 is obtained for block B. (In some examples, passage 138a does not use neural network, but other techniques known in the art).

At this point, the image im1 in block B (138) has size MxN, while the image to be displayed is requested to have size M1xN1. It has been noted, however, that it is simply possible to perform a transformation (e.g., U) 140 which transports the image im1 in block B (138) into M1xN1.

Where D as performed at 134 is a downsampling operation, it has been noted that U at 140 may be an upsampling operation. Therefore, U (140) may be obtained by introducing

coefficients in the M1xN1 block, besides the coefficients in the MxN block 138 as obtained at operation 138a with neural network.

For example, in case of M1=2M and N1=2N, it is simply possible to perform an interpola-tion (e.g., bilinear interpolation), so as to approximate (“guess”) the coefficients of im1 that had been discarded by the transformation D. An M1xN1 image im1 is therefore obtained as element 142, and may be used for displaying the block image as part of the image 10.

Notably, it is also theoretically possible to obtain the block 144, which, notwithstanding, would be the same of the template 60 (apart from errors due to the transformations D and U). Therefore, advantageously, there is no necessity of transforming Brec for obtaining a new version of B1,rec which is already at disposal as the template 60.

Operations shown in Fig. 13a may be performed at the predictor 44 or 44’, for example. The M1xN1 image im1 (142) may therefore be understood as the prediction signal 24 (Fig.

2) or 24’ (Fig. 4) to be summed with the prediction residual signal output by the inverse transformer 40 or 40' for obtaining the reconstructed signal.

Fig. 13b shows an example, in the transform domain (e.g., in the examples of Figs. 11-1, 11-2). An element 162 is represented as formed by the spatial-domain template 60 (al-ready decoded) and the spatial-domain block 18 (with unknown coefficients). Block 18 may have a size M1xN1 and may have unknown coefficients, which are to be determined, for example, at the predictor 44-1 or 44-2.

There arises the possibility that, while having at disposal neural networks for a determined MxN size, there are no neural networks to directly operate with M1xN1 blocks in the trans-form domain.

However, it has been noted that it is possible to use, at the predictor 44-1, 44-2, a trans-formation D (166) applied to the template 60 (“second template”) to obtain a spatial-domain template 170 with different dimensions (e.g., reduced dimensions). The template

170 (“first template”) may have an L-formed shape, e.g., such as the shape of the tem-plate 130 (see above).

At this point, at passage 170a, the neural networks (e.g., 800-80N) may be applied accord-ing to any of the examples above (see Figs. 5-10). Hence, at the end of passage 170a, the known coefficients for a version 172 of block 18 may be obtained.

It is noted, however, that the dimensions MxN of 172 do not fit the dimensions M1xN1 of the block 18 which has to be visualized. Hence, a transformation (e.g., at 180) into trans-form domain may be operated. For example, an MxN transform-domain block T (176) may be obtained. In order to increase the number of the rows and the columns to M1 and N1, respectively, a technique known as zero padding may be used, e.g., by introducing values "0" in correspondence to frequency values associated to frequencies which do not exist in the MxN transform T (176). A zero-padding area 178 may therefore be used (it may have an L shape, for example). Notably, the zero-padding area 178 comprises a plurality of bins (all zero) which are inserted to the block 176 to obtain the block 182. This may be ob-tained with a transformation V from T (transformed from 172) to T1 (182). While the di-mensions of T (176) do not conform with the dimensions of the block 18, the dimensions of T1 (182), by virtue of the insertion of the zero-padding area 178, actually conform to the dimensions of the block 18. Furthermore, the zero-padding is obtained by inserting higher-frequency bins (which have a zero value), which has a result analogous to an interpola-tion.

Hence, at adder 42-1, 42-2, it is possible to add the transform T1 (182), which is a version of 45-1 , 45-2. Subsequently, the inverse transformation T-1 may be performed to obtain the reconstructed value 60 in the spatial domain to be used for visualizing the picture 10.

The encoder may encode in the datastream 12 information regarding the resampling (and the use of neural networks for blocks with different size from that of the block 18), so that the decoder has the knowledge thereof.

Discussion

Let B1 (e.g., 18) be a block with M1 rows and N1 columns and assume that M1 ≥ M and N1 ≥ N. Let B1, rec be a neighborhood of (e.g., template 60 in adjacent to) B1 and assume that the region Brec (e.g., 130) is regarded as a subset of B1,rec. Let im1 (e.g., 138) an im-age on B1 and let reci (e.g., the coefficients on Bi,rec) be an already reconstructed image on B1,rec. The solutions above are based on a predefined downsampling operation D (e.g., 134, 166) which maps images on B1, rec to images on B1. For example, if M1 = 2M, N1 =

2N, if Brec consists of K rows above B and L columns left of B and a corner of size K × L on the top left of B and if B1, rec consists of 2K rows above B1 and 2L columns left of B and a corner of size 2K × 2L on the top left of B1, then D can be the operation of applying a smoothing filter followed by a factor two downsampling operation in each direction.

Thus, D(rec1) can be regarded as a reconstructed image on Brec. Using our Neural-Network-based intra prediction modes as above, out of D(rec1) we can form the prediction signal pred (45-1) which is an image on B.

Now we differentiate two cases: First, we assume that, on B, our Neural-Network-based intra prediction predicts into the sample (spatial) domain, as in Figs. 2, 4, and 13a. Let U (140) be a fixed upsampling filter that maps images on B (e.g., 138) to images on B1 (e.g., 142). For example, if M1 = 2M and N1 = 2N, then U can be the bilinear interpolation opera-tion. Then we can form U(pred ) to obtain an image on B1 (e.g., 45-1) that we regard as a prediction signal for im1 (e.g., 10)

Second, we assume that, as in Figs. 11-1, 11-2, and 13b, on B the prediction signal pred

(e.g., 45-2) is to be regarded as a prediction signal in the transform domain with respect to an image transform T on B with inverse transform S. Let T1 be an image transform on B1 with inverse transform S1. Assume that a predefined mapping V is given which maps im-ages from the transform domain of T to the transform domain of T1. If for example T is the Discrete Cosine Transform on M×N blocks with inverse transform S, and Ti is the Discrete Cosine Transform on M1 × N1 with inverse transform Si, then a block of transform coeffi-cients on B can be mapped to a block of transform coefficients on B1 by zero padding and scaling (see, for example, 178). This means that one sets all transform coefficients on B1 to zero if the position in the frequency space is larger than M or N in the horizontal resp. vertical direction and that one copies the appropriately scaled transform coefficients on B to the remaining M * N transform coefficients on B1. Then we can form V (pred) to obtain an element of the transform domain for T1 that is to be regarded as a prediction signal for T1 (im1). The signal V (pred) might now be further processed as described above.

As explained above in respect to Figs. 1-10, we also described how a Neural Network based operation can be used to rank several intra-prediction modes on a given block B by generating a conditional probability distribution among these modes and how this ranking can be used to signal which intra prediction mode is to be applied at the current block. Using a downsampling operation (e.g., 166) on the input of the Neural Network that gen-erates the latter ranking in the same way as for the actual prediction modes yields a rank-ing for the extension of the prediction modes to the larger Block B1 just described and can thus be used to signal which extended mode is to be used on the block B1. Whether on a given block B1 Neural Network based intra prediction modes from a smaller block B are to be used to generate a prediction signal can be either predefined or can be signaled as a side information in the underlying video codec.

Other examples

In general terms, a decoder as above maybe and/or comprise an encoder as above or vice versa. For example, the encoder 14 may be or comprise the decoder 54 (or vice ver-sa); encoder 14-1 may be the decoder 54-2 (or vice versa), etc. Further, the encoder 14 or 14-1 may also be understood as containing itself a decoder, as the quantized prediction residual signals 34 form a stream which is decoded to obtain the prediction signal 24 or 24-1.

Although some aspects have been described in the context of an apparatus, it is clear that these aspects also represent a description of the corresponding method, where a block or device corresponds to a method step or a feature of a method step. Analogously, aspects described in the context of a method step also represent a description of a corresponding block or item or feature of a corresponding apparatus. Some or all of the method steps may be executed by (or using) a hardware apparatus, like for example, a microprocessor, a programmable computer or an electronic circuit. In some examples, one or more of the most important method steps may be executed by such an apparatus.

The inventive encoded data stream can be stored on a digital storage medium or can be transmited on a transmission medium such as a wireless transmission medium or a wired transmission medium such as the Internet.

Depending on certain implementation requirements, examples of the invention can be implemented in hardware or in software. The implementation can be performed using a digital storage medium, for example a floppy disk, a DVD, a Blu-Ray, a CD, a ROM, a PROM, an EPROM, an EEPROM or a FLASH memory, having electronically readable control signals stored thereon, which cooperate (or are capable of cooperating) with a programmable computer system such that the respective method is performed. Therefore, the digital storage medium may be computer readable.

Some examples according to the invention comprise a data carrier having electronically readable control signals, which are capable of cooperating with a programmable computer system, such that one of the methods described herein is performed.

Generally, examples of the present invention can be implemented as a computer program product with a program code, the program code being operative for performing one of the methods when the computer program product runs on a computer. The program code may for example be stored on a machine readable carrier.

Other examples comprise the computer program for performing one of the methods de-scribed herein, stored on a machine readable carrier.

In other words, an example of the inventive method is, therefore, a computer program having a program code for performing one of the methods described herein, when the computer program runs on a computer.

A further example of the inventive methods is, therefore, a data carrier (or a digital storage medium, or a computer-readable medium) comprising, recorded thereon, the computer program for performing one of the methods described herein. The data carrier, the digital storage medium or the recorded medium are typically tangible and/or non‒transitionary.

A further example of the inventive method is, therefore, a data stream or a sequence of signals representing the computer program for performing one of the methods described herein. The data stream or the sequence of signals may for example be configured to be transferred via a data communication connection, for example via the Internet.

A further example comprises a processing means, for example a computer, or a pro-grammable logic device, configured to or adapted to perform one of the methods de-scribed herein.

A further example comprises a computer having installed thereon the computer program for performing one of the methods described herein.

A further example according to the invention comprises an apparatus or a system config-ured to transfer (for example, electronically or optically) a computer program for perform-ing one of the methods described herein to a receiver. The receiver may, for example, be

a computer, a mobile device, a memory device or the like. The apparatus or system may, for example, comprise a file server for transferring the computer program to the receiver.

In some examples, a programmable logic device (for example a field programmable gate array) may be used to perform some or all of the functionalities of the methods described herein. In some examples, a field programmable gate array may cooperate with a micro-processor in order to perform one of the methods described herein. Generally, the meth-ods are preferably performed by any hardware apparatus.

The apparatus described herein may be implemented using a hardware apparatus, or using a computer, or using a combination of a hardware apparatus and a computer.

The apparatus described herein, or any components of the apparatus described herein, may be implemented at least partially in hardware and/or in software.

The methods described herein may be performed using a hardware apparatus, or using a computer, or using a combination of a hardware apparatus and a computer.

The methods described herein, or any components of the apparatus described herein, may be performed at least partially by hardware and/or by software.

The above described examples are merely illustrative for the principles of the present in-vention. It is understood that modifications and variations of the arrangements and the details described herein will be apparent to others skilled In the art. It is the intent, there-fore, to be limited only by the scope of the impending patent claims and not by the specific details presented by way of description and explanation of the examples herein.
Claims

1. Apparatus (54-2) for block-wise decoding a picture (10) from a data stream (12), the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal for a block (136, 172) of a predetermined size of the picture is deter-mined by applying a first template (130, 170) of samples which neighbours the current block onto a neural network (80), wherein the apparatus is configured, for a current block (18) differing from the predetermined size, to:

resample (134, 166) a second template (60) of samples neighboring the current block (18), so as to conform with the first template (130, 170) so as to obtain a resampled template (130, 170);

apply (138a, 170a, 44-1, 44-2) the resampled template (130, 170) of samples onto the neural network (80) so as to obtain a preliminary intra-prediction signal (138, 172, 176); and

resample (140, 180) the preliminary intra-prediction signal (138, 172, 176) so as to conform with the current block (18) so as to obtain (140) the intra-prediction signal (142, 24-1 , 24-2) for the current block (18).

2. Apparatus of claim 1, configured to resample (134, 166) by downsampling (D) the second template (60) to obtain the first template (130, 170).

3. Apparatus of claim 1 or 2, configured to resample (140) the preliminary intra-prediction signal (138, 172, 176) by upsampling (U, 140) the preliminary intra-prediction signal (138, 172).

4. Apparatus of any of the preceding claims, configured to:

transform the preliminary intra-prediction signal (138, 176) from a spatial domain into a transform domain; and

resample the preliminary intra-prediction signal (176) in the transform domain.

5. Apparatus of claim 4, configured to:

resample (140) the transform-domain preliminary intra-prediction signal (176) by scaling the coefficients of the preliminary intra-prediction signal (176).

6. Apparatus of claim 4 or 5, configured to:

resample (140) the transform-domain preliminary intra-prediction signal (176) by: increasing the dimensions of the intra-prediction signal (176) to conform (182) to the dimensions of the current block (18); and

zero-padding the coefficients of added coefficients of the preliminary intra- prediction signal (176), the added coefficients relating to higher-frequencies bins.

7. Apparatus of any of claims 4-6, configured to:

compose (42-1, 42-2) the transform-domain preliminary intra-prediction signal

(176) with a dequantized version (39-1 , 39-2) of a prediction residual signal (34).

8. Apparatus of any of the preceding claims, configured to:

resample (140) the preliminary intra-prediction signal (138) in the spatial domain.

9. Apparatus of claim 8, configured to:

resample (140) the preliminary intra-prediction signal (138) by performing a bilinear interpolation.

10. Apparatus of any of the preceding claims, further configured to:

encode in a data field information regarding the resampling and/or the use of neu-ral networks for different dimensions.

11. Apparatus of any of the preceding claims, the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal (24-1 , 24-2) for a cur-rent block (18) of the picture is determined by:

applying a first set (60) of neighboring samples of the current block (18) onto a neural network (80) to obtain a prediction (45-1, 45-2) of a set of transform coefficients of a transform (34, 39-1 , 39-2) of the current block (18).

12. Apparatus for block-wise decoding a picture (10) from a data stream (12), the ap-paratus supporting at least one intra-prediction mode according to which the intra-prediction signal (24-1, 24-2) for a current block (18) of the picture is determined by:

applying a first set (60) of neighboring samples of the current block (18) onto a neural network (80) to obtain a prediction (45-1, 45-2) of a set of transform coefficients of a transform (34, 39-1 , 39-2) of the current block (18).

13. Apparatus of claim 12, further configured as an apparatus of any of claims 1-11,

14. Apparatus of any of the preceding claims, configured to:

inversely transform (40-1, 40-2) the prediction (45-1, 45-2) to obtain a reconstruct-ed signal (24-1, 24-2).

15. Apparatus of any of the preceding claims, configured to:

decode from the data stream (12) an index (70b) using a variable length code; and perform the selection using the index (70b).

16. Apparatus of any of the preceding claims, configured to:

determine a ranking of the set (72) of intra prediction modes; and, subsequently, resample the second template (60).

17. Apparatus (14-1) for block-wise encoding a picture (10) into a data stream (12), the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal for a block (136, 172) of a predetermined size of the picture is deter-mined by applying a first template (130, 170) of samples which neighbours the current block onto a neural network (80), wherein the apparatus is configured, for a current block (18) differing from the predetermined size, to:

resample (134, 166) a second template (60) of samples neighboring the current block (18), so as to conform with the first template (130, 170) so as to obtain a resampled template (130, 170);

apply (138a, 170a, 44-1, 44-2) the resampled template (130, 170) of samples onto the neural network (80) so as to obtain a preliminary intra-prediction signal (138, 172,

176); and

resample (140, 180) the preliminary intra-prediction signal (138, 172, 176) so as to conform with the current block (18) so as to obtain (140) the intra-prediction signal (142, 24-1, 24-2) for the current block (18).

18. Apparatus of claim 17, configured to resample (134, 166) by downsampling (D) the second template (60) to obtain the first template (130, 170).

19. Apparatus of claim 17 or 18, configured to resample (140) the preliminary intra-prediction signal (138, 172, 176) by upsampling (U, 140) the preliminary intra-prediction signal (138, 172).

20. Apparatus of any of the preceding claims, configured to:

transform the preliminary intra-prediction signal (138, 176) from a spatial domain into a transform domain; and

resample the preliminary intra-prediction signal (176) in the transform domain.

21. Apparatus of claim 20, configured to:

resample (140) the transform-domain preliminary intra-prediction signal (176) by scaling the coefficients of the preliminary intra-prediction signal (176).

22. Apparatus of claim 20 or 21 , configured to:

resample (140) the transform-domain preliminary intra-prediction signal (176) by:

increasing the dimensions of the intra-prediction signal (176) to conform (182) to the dimensions of the current block (18); and

zero-padding the coefficients of added coefficients of the preliminary intra- prediction signal (176), the added coefficients relating to higher-frequencies bins.

23. Apparatus of any of claims 20-22, configured to:

compose (42-1, 42-2) the transform-domain preliminary intra-prediction signal (176) with a dequantized version (39-1, 39-2) of a prediction residual signal (34).

24. Apparatus of any of claims 17-23, configured to:

resample (140) the preliminary intra-prediction signal (138) in the spatial domain.

25. Apparatus of claim 24, configured to:

resample (140) the preliminary intra-prediction signal (138) by performing a bilinear interpolation.

26. Apparatus of any of claims 17-23, further configured to:

encode in a data field information regarding the resampling and/or the use of neu-ral networks for different dimensions.

27. Apparatus of any of claims 17-26, the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal (24-1, 24-2) for a current block (18) of the picture is determined by:

applying a first set (60) of neighboring samples of the current block (18) onto a neural network (80) to obtain a prediction (45-1, 45-2) of a set of transform coefficients of a transform (34, 39-1, 39-2) of the current block (18).

28. Apparatus for block-wise decoding a picture (10) from a data stream (12) and/or for block-wise encoding a picture (10) into a data stream (12), the apparatus supporting at least one intra-prediction mode according to which the intra-prediction signal (24-1, 24-2) for a current block (18) of the picture is determined by:

applying a first set (60) of neighboring samples of the current block (18) onto a neural network (80) to obtain a prediction (45-1, 45-2) of a set of transform coefficients of a transform (34, 39-1, 39-2) of the current block (18).

29. Apparatus of claim 28, further configured as an apparatus of any of claims 17-28.

30. Apparatus of any of claims 17-29, configured to:

inversely transform (40-1, 40-2) the prediction (45-1 , 45-2) to obtain a reconstruct-ed signal (24-1, 24-2).

31. Apparatus of any of claims 17-30, configured to:

decode from the data stream (12) an index (70b) using a variable length code; and perform the selection using the index (70b).

32. Apparatus of any of claims 17-31, configured to:

determine a ranking of the set (72) of intra prediction modes; and, subsequently, resample the second template (60).

33. Method for block-wise decoding a picture (10) from a data stream (12), the method supporting at least one intra-prediction mode according to which the intra-prediction signal for a block (136, 172) of a predetermined size of the picture is determined by applying a first template (130, 170) of samples which neighbours the current block onto a neural net-work (80), the method comprising:

resampling (134, 166) a second template (60) of samples neighboring the current block (18), so as to conform with a first template (130, 170) so as to obtain a resampled template (130, 170),

applying (138a, 170a, 44-1, 44-2) the resampled template (60) of samples onto a neural network (80) so as to obtain a preliminary intra-prediction signal (138, 172, 176), and

resampling (140) the preliminary intra-prediction signal (138, 172, 176) so as to conform with the current block (18) so as to obtain (140) the intra-prediction signal (142, 24-1, 24-2) for the current block (18).

34. Method for block-wise encoding a picture (10) into a data stream (12), the method supporting at least one intra-prediction mode according to which the intra-prediction signal for a block (136, 172) of a predetermined size of the picture is determined by applying a first template (130, 170) of samples which neighbours the current block onto a neural net-work (80), the method comprising:

resampling (134, 166) a second template (60) of samples neighboring the current block (18), so as to conform with a first template (130, 170) so as to obtain a resampled template (130, 170),

applying (138a, 170a, 44-1, 44-2) the resampled template (60) of samples onto a neural network (80) so as to obtain a preliminary intra-prediction signal (138, 172, 176), and

resampling (140) the preliminary intra-prediction signal (138, 172, 176) so as to conform with the current block (18) so as to obtain (140) the intra-prediction signal (142, 24-1, 24-2) for the current block (18).

35. Method for block-wise decoding a picture (10) from a data stream (12), comprising:

applying a first set (60) of neighboring samples of a current block (18) onto a neu-ral network (80) to obtain a prediction (45-1, 45-2) of a set of transform coefficients of a transform (34, 39-1, 39-2) of a current block (18).

36. Method for block-wise encoding a picture (10) into a data stream (12), comprising:

applying a first set (60) of neighboring samples of a current block (18) onto a neu-ral network (80) to obtain a prediction (45-1, 45-2) of a set of transform coefficients of a transform (34, 39-1, 39-2) of a current block (18).

37. A computer-readable storage medium comprising instructions which, when exe-cuted by a computer, cause the computer to perform a method according to any of claims 33-36.

38. A data stream encoding a picture (10) and obtained by a method supporting at least one intra-prediction mode according to which the intra-prediction signal for a block (136, 172) of a predetermined size of the picture is determined by applying a first template (130, 170) of samples which neighbours the current block onto a neural network (80), the method comprising:

resampling (134, 166) a second template (60) of samples neighboring the current block (18), so as to conform with a first template (130, 170) so as to obtain a resampled template (130, 170),

applying (138a, 170a, 44-1, 44-2) the resampled template (60) of samples onto a neural network (80) so as to obtain a preliminary intra-prediction signal (138, 172, 176), and

resampling (140) the preliminary intra-prediction signal (138, 172, 176) so as to conform with the current block (18) so as to obtain (140) the intra-prediction signal (142, 24-1, 24-2) for the current block (18).

39. A data stream encoding a picture (10) and obtained by a method comprising: applying a first set (60) of neighboring samples of a current block (18) onto a neu-ral network (80) to obtain a prediction (45-1 , 45-2) of a set of transform coefficients of a transform (34, 39-1, 39-2) of a current block (18).