a Method and system for digitally enhancing image
resolution and quality of image sequence data
related applications
This application claims priority to U.S. Provisional Application No.60/365,021 filed
March 13, 2002 entitled, "Systems and Methods for Digitally Re-Mastering or Otherwise
Modifying Motion Pictures or Other Image Sequences Data for Alternative Formate including
Large Format Projection or Other Purposes", which is incorporated by reference herein.
Field of the Invention
The present invention is related to a method and system for digitally enhancing
image resolution and quality of image sequence data and, more specifically to a system anc
a method that efficiently computes image data from an entire motion picture in a parallel
and pipeline fashion for the purpose of concurrent release with the original format, and to a
method that enhances images by improving image resolution and quality for exhibition,
typically in an alternative format including large format projection environment. The present
invention is also applicable to the enhancement of a broad range of image sequences
originated from film, video, optical devices, electronic sensors, and so on. It additionally
may be employed to improve quality of images for display in their original format.
Background of the Invention
A motion picture produced in a 35mm film format is intended to be exhibited in a
conventional format cinema or in other smaller formats like home video and broadcast
television. The display resolution required to maintain adequate display quality can be
calculated based on the screen size, the theatre geometry, audience seating positions as well
as the minimum visual acuity that needs to be maintained in order to deliver the required
image quality. In a conventional cinema, a display resolution of around 2000 pixels across the
width of the screen is considered adequate for delivery of satisfactory image quality. This
resolution requirement is largely supported by the 35mm film formats as well as by existing
film production process chain from original photography, post-production, to film laboratory
process. A similar display resolution requirement is also recommended for digital cinemas
designed to replace conventional film-based cinemas.
In a large format cinematic venue, audiences expect a significantly superior visual
experience to that which they perceive in a conventional cinema. Audiences in a large format
cinema enjoys a field of view much larger than that from a conventional cinema. To maintain
a superior visual experience in a large format cinema, the film production chain must deliver
a spatial image resolution much higher than that for a conventional cinema. The current 15/70
film format production process chain adequately supports this higher spatial resolution
requirement. However, when a motion picture originated for the 35mm film format is to be
exhibited in a large format cinema venue, the existing production system and process cannot
deliver sufficient image quality. The present invention defines a method of digitally re-
mastering a 35mm motion picture with enhanced image resolution and quality as demanded
by the large format cinematic experience.
The digital re-mastering process of the present invention is primarily (but not
exclusively) used for the enhancement of image resolution of a motion image sequence
originating with live action film photography. This process can be applied to the
enhancement of computer generated animation or cell animation images that have been
transferred to film. The digital re-mastering process can also be used to enhance the
resolution and the quality of moving images captured using an optical imaging device or an
electronic sensor device.
One aspect of the digital re-mastering concept of the present invention is that spatial
resolution of each image frame in a live action motion picture sequence can be enhanced
through temporal image processing. This is very different from the concept of film restoration
in which "cleaning up" noise and "removing" artifacts are primary goals. There were many
successful film restoration projects in the past two decades, and typical examples are
Disney's Snow White and Seven Dwarves re-release and subsequent George Lucas1 Star
Wars trilogy re-releases. Most film restoration methods are designed to compensate for the
loss of image quality caused by the deterioration of film conditions and to restore the
characteristics of images close to the original form. Since the targeted re-release platform for
a film restoration project is usually the same conventional cinema that the film was originally
intended to be exhibited, or even smaller exhibition formats like home video and television,
enhancing the spatial resolution of original imagery is not a major concern for film
restoration.
The digital re-mastering process of the present invention should also be distinguished
from existing methods for re-mastering an animated motion picture for large format releases,
such as Disney's Fantasia 2000 and Beauty and Beast large format release production. In
those efforts, image data was originally created in a digital form and was not corrupted by a
film transfer process. As a result, the spatial resolution of image frames cannot be further
enhanced through image processing methods unless those images are re-rendered in more
detail. The method used in Fantasia 2000 and Beauty and Beast re-releases cannot enhance
the image resolution of live action film photography.
The most straightforward method of displaying a 35mm film originated motion
picture in a large format cinema is to use a projection lens with a larger magnification to fill
the entire screen. This method cannot deliver sufficient visual quality due to the fact that
images on a 35mm release print do not have sufficient spatial resolution. A better method is
to digitally enlarge each frame of the motion picture using digital spatial interpolation
methods and record the enlarged image data onto a large format film, like the 15/70 film
format, for projection. The existing spatial interpolation methods do not improve spatial
resolution and often soften images. Certain spatial high-pass filtering methods can be used to
improve perceived image sharpness, but those methods also emphasize the noise in the
images, like film grain. To reduce image noise, certain low-pass spatial filters can be
applied, but those filters inevitably remove image details. Therefore, conventional spatial
processing methods cannot satisfy conflicting demands for noise reduction and maintaining
image sharpness.
Summary of the Invention
The image re-mastering method of the present invention provides a solution to
achieve both resolution enhancement and noise reduction. This method states that image
spatial resolution can be enhanced through temporal image processing. For this method to be
applied in the most effective way, the process requires that all image details on its original
form be preserved. For images originated on film, the basic elements are film grains. It is
well known, as stated by the Sampling Theorem, that all information can be preserved if the
spatial sampling grid satisfies the Nyquist sampling frequency, which is twice that of the
spatial frequency limit of the image content. When scanning a camera negative film, this is
equivalent to using a pixel pitch of no larger than 6 ten, in order to capture image details
down to film grain level. For an image frame on a 35mm film photographed with Academy
aperture of 0.825"x0.602", a scanning resolution of at least 3500x2550 pixels is required.
Spatial resolution enhancement is more effective when each image frame is first
digitally enlarged using spatial interpolation. Spatial interpolation does not improve spatial
resolution, but it expands the image frequency spectrum in the frequency domain so that extra
room is available for the additional high-frequency details to be added to the images. This
resolution enhancement concept is shown in Figure 1. The additional image details can be
recovered through a temporal filtering process. In a motion sequence, an object in a scene is
captured on a sequence of frames, and each frame contains a similar but not identical version
of the object. It is possible that certain image details about the object that are not explicit on
the current frame may be captured on its neighboring frames. By collecting all information
about the object from neighboring frames to improve the current frame, the resulting object
on the current frame may exhibit improved details that do not exit in its original form. This
concept can be realized through a temporal filtering method in which information from a
number of frames is analyzed and combined so that additional image details may be
recovered for every frame in an motion sequence. A variation of the temporal filtering
method can also be used to reduce temporally uncorrelated noise, like perceived film
graininess in an image sequence. A new method for improving image sharpness through the
enhancement of MTF measurement is also described.
The temporal processing methods require computationally expensive motion
estimation algorithms. One of the challenges is to develop a computing system that
implements the present invention in a highly efficient way so that re-mastering a complete
motion picture can be achieved in a relatively short period of time. The present invention
describes a parallel and distributed computing system with an intelligent central control that
manages the image render process with maximum efficiency. The intelligent central control
automates the image data enhancement process using various types of optimization schemes.
The computing system has a unique architecture that is scalable to any throughput
requirement.
Another challenge of the implementation is to define a process that provides
functionality to meet a wide range of requirements from a high-quality motion picture re-
mastering production. In the present invention, an entire process of re-mastering operations is
described that includes various stages and process modules. The core part of the process can
be implemented in an automated mode, but it also allows maximum flexibility for human
user input and interaction. A set of original algorithms for the estimation of optimal
parameters and for automated quality evaluation based on statistical analysis are described in
the present invention, and those algorithms are managed by the intelligent central control to
deliver maximum quality results.
The process and system described in the present invention is designed to meet the
most rigorous production demands, including a concurrent release of a re-mastered format of
a new motion picture with its original rele^e. This application requires the re-mastering
process and system to be able to accurately and reliably track the status of every operation in
the pipeline as well as the status of data flow of every piece of image data. The system
configuration described in this invention allows the intelligent central control to track the
status of every device throughout the entire process, including facilities remotely located. The
intelligent central control also provides up-to-date status reports and answers user specified
queries. Those features have proved to be extremely important for meeting a rigorous
production schedule for a motion picture re-mastering project
The teachings of the present invention can be readily understood by considering the
following detailed descriptions in conjunction with the accompanying drawings, in which:
Figure 1 depicts the concept of spatial resolution enhancement represented in the
frequency domain;
Figure 2 is a process flow chart describing a digital re-mastering process for motion
picture and other moving images;
Figure 3 depicts the system configuration of the Data Processing stage of the image
enhancement system for motion picture re-mastering;
Figure 4 is a process flow chart of the Pre-processing module;
Figure 5 is a process flow chart of the Render module;
Figure 6 describes the algorithm for automatic prediction of render parameters;
Figure 7 describes the process pipeline of a render client;
Figure 8 depicts a typical temporal filtering scheme with a temporal window;
Figure 9 describes three temporal filtering schemes for noise reduction;
Figure 10 depicts three schemes for render job distribution;
Figure 11 is a process flow chart of the Verification module;
Figure 12 is a process flow chart of the Post-processing module; and
Figure 13 shows an example of the types of production information that are tracked
by the Intelligent Controller.
The present invention describes a process and methods for digitally re-mastering a
motion picture or any moving picture sequence from its original format to an alternative
format or its original format with enhanced image quality, as well as a system
implementation of that process. The digital image re-mastering process is depicted in Figure
2, and the system implementation of the process is depicted in Figure 3. As depicted in Figure
2, the digital image re-mastering process consists of four stages: Format Conversion 100,
Data Processing 110, Image Out 120 and Image Approval 130. The process is controlled by
a central control system 140, and at the heart of the central control is the Intelligent
Controller 141. The Intelligent Controller 141 is implemented through a combination of
hardware and software, and it controls and monitors every aspect of the process from tracking
physical data flow to controlling actual task execution.
At the first stage Format Conversion 100, images of a motion picture or any moving
images are converted to a digital format that can be handled by the Data Processing 110
stage. The majority of motion pictures originate in film (negative film or intermediate film),
and need to be converted to digital format through a device called a film scanner (Glenn
Kennel, "Digital Film Scanning and Recording: The Technology and Practice", SMPTE
Journal, Vol. 103, No. 3, March 1994). A film scanner is a sampling device, and it converts
each image sample from film density to binary digits representing red, green and blue (RGB)
color components. The scanned data needs to have a sufficient bit-depth in order to preserve
the full dynamic range of the images on film. In one embodiment of the invention, each
image frame is sampled with a pixel pitch no greater than 6 ym, and each sample is quantized
into RGB channels with 10-bit bit-depth for each channel. The scanned data is then packed in
a file format called DPX (Digital Moving Picture Exchange) format (SMPTE 268M) or its
earlier version named Cineon format. For images originated in a non-digital form other than
film, the digitization process needs to support a sampling rate and a quantization level that
preserves all the information of the original images. As motion pictures are increasingly
produced and released in digital format, the available digital data can be directly converted
from their original format to any other format through a digital conversion process 102. In
one embodiment of the invention, the desirable format is the 10-bit DPX format or Cineon
format The format conversion process 102 typically comprises operations such as color
space conversion, compression/decompression and bit-packing, as one skilled in the art will
readily be able to adapt the process described by the teaching to any data formats. The
converted image data must be visually inspected as correct through a Data Inspection process
103 so that errors in image data caused by incorrect setting of film scanners and other failures
in the conversion process are identified. Another function of the Data Inspection 103 is to
ensure that the converted image data meets a set of pre-defined quality standards including
the preservation of dynamic range and image details. Image data that contains errors or does
not meet quality standards is rejected to be re-scanned or re-converted.
The image data converted at the Format Conversion stage 100 is referred to as
Original Image Data. At the next Data Processing stage 110, the Original Image Data is
enhanced both in resolution and visual quality through an image enhancement process The
image enhancement process comprises four process modules, and they are the Pre-processing
module 111, the Render module 112, the Verification module 113 and the Post-processing
module 114.
The Pre-processing module 111 comprises processes that depend on human user's
decisions, such as shot separation, re-framing, color correction and scene classification. It
also includes a process for identifying and removing artifacts existing in the Original Image
Data. In one embodiment of the invention, all operations required for decision making are
implemented by a combination of special-purpose software and user workstations that allow
user interactions. The user's decisions are collected by the Intelligent Controller 141 that then
instructs user's workstations to apply corresponding pre-processing operations to the Original
Image Data. In another embodiment of the present invention, the actual pre-processing
operations are implemented in the next Render module as part of render client process
pipeline, and those operations are totally controlled by the Intelligent Controller 141 based on
user's decisions. In another embodiment of the invention, decisions on color correction and
artifact identification are made by image analysis algorithms so that those operations can be
implemented in a fully automated mode without need for human intervention.
The Render module 112 is responsible for the enhancement of image data in both
resolution and visual quality. The system implementation of the Render module 112 is a
combination of hardware and software designed for achieving high throughput and high
computational efficiency. In one embodiment of the present invention, the Render module
mainly performs three types of enhancement operations to the image data: resolution-
enhancement, noise reduction and sharpening. In another embodiment of the present
invention, the Render module also performs additional pre-processing operations including
color correction, re-framing and artifact removal. To achieve high throughput, the Render
system implementation adopts a parallel computing architecture with multiple computing
render clients controlled by the Intelligent Controller 141.
The Intelligent Controller 141 is responsible for maximizing render efficiency and
delivering optimal image quality. It achieves this by using intelligent algorithms to compute
the best solutions. In one embodiment of the invention, the operation of the Render module
112 is fully automated. The Intelligent Controller determines how image data is distributed
among available render clients to achieve maximum efficiency. It also determines the best
set of render instructions to be used to achieve optimal render quality. The Intelligent
Controller constantly monitors the performance of the render operations and modifies its
instructions. In the same embodiment of the invention, users are permitted to input their
preference to the Intelligent Controller to be used for making render decisions and even
overwrite decisions made by automated schemes. In another embodiment of the invention,
users issue render instructions based on their experience and observation and constantly make
changes to the instructions based on statistical performance analysis. In both embodiments of
this invention, the quality of the enhanced image data is to be visually verified in the
Verification module 113 to ensure they meet pre-defined quality and resolution criteria.
In the Verification module 113, the enhanced image data that does not meet the
quality and resolution criteria is rejected and sent back to the Render module 112 to be
rendered with a modified set of instructions or to be sent further back to the Pre-processing
module 111 if problems are related to Pre-processing decisions. In one embodiment of the
present invention, the Intelligent Controller 141 makes decisions on how to modify render
instructions based on some statistical quality indicators calculated in the Render module 112.
In another embodiment of the invention, users make modification decisions with the
assistance of those statistical quality indicators. In general cases, image data does not pass the
Verification module until it meets the pre-defined quality and resolution criteria. In special
cases where image data contains certain artifacts that can be fixed at the Post-processing 114
module, the image data will be sent to the Post-processing module 114 for fixing before being
sent back to the Verification module 113 for final evaluation.
The Post-processing module 114 performs certain final problem-fixing operations.
The image data that meets all quality standards is organized before being sent to the Image
Out stage 120. In certain cases, image data may need to be converted to the format specified
by the film out process 121 or the digital out process 122 . The enhanced image data is also
written to permanent data storage, such as data tapes or optical discs, for data transfer or
backup purposes.
Image Out 120 is a stage where enhanced image data of a motion picture is recorded
onto an alternative film format, or re-formatted for digital display. In the case of film output
121, film recorders are used to transfer image data onto film, and the recorded film is
processed with a standard laboratory process. A print film with correct timing is made using
an optical printer. In the case of digital output 122, image data must be converted to the
format suitable for digital display, and the operations involved in the conversion process may
require resizing, color space conversion, re-framing and compression. Those skilled in the art
will readily be able to convert the data described by the teaching to any output data formats.
The final stage is Image Approval 130, and it is performed by human inspection of the
motion picture in an intended release format. In the case of film format, the print film is
projected in a viewing environment representative to the cinemas where the motion picture is
planned for release. The approval process can be done in segments or in the entirety of the
motion picture. Those image segments that are rejected will be sent to appropriate earlier
stages of the process to be re-processed. In the case of digital display, a similar screening
process is performed using a representative digital display system. The approved images
become the final re-mastered image product.
In the present invention, the physical facilities needed for each stage of the digital re-
mastering process may not need to be located in proximity to each other. In one embodiment
of the invention, the facility for each stage is located in a separate geographical location. The
exchange of image data and other information between different locations can easily be done
by electronic data transmission and through a courier service.
Figure 3 depicts the system configuration of the Data Processing stage 110
implemented in a single location. It can be divided into two functional components. The first
component is the Operation System 150, which supports all functions in Pre-processing 111,
Verification 113 and Post-processing 114 modules. The Operation System 150 consists of a
Data File Server 151, a Central Data Storage 152 with sufficient disk capacity, multiple
workstations 153 equipped with special-purpose software tools for user interactive
operations, a network switch that provides high-bandwidth connections between the file
server and workstations, and a number of tape drives 155 that serve as data input and output
devices.
The second component is the Render System 160, which supports all operations in the
Render module 112. The Render System 160 consists of an Intelligent Controller Server 161,
a Controller Data Storage 162 supporting the server, and multiple render client computing
devices 163 configured in clusters. Given the processing time needed for each image frame,
the number of render clients allowed for each cluster is limited by the bandwidth of the
cluster network switch 165. A backbone network switch 164 provides high-bandwidth
connections to all cluster network switches 165, and the number of clusters supported by the
system is limited by the bandwidth of the backbone switch 164. The data transfer between the
Operation System 150 and the Render System 160 is through a high bandwidth link between
the Data File Server 151 and the Intelligent Controller Server 161. The system described in
Figure 3 has a modular design and is totally scalable to daily throughput requirement from a
specific re-mastering project.
In one embodiment of the present invention, the Intelligent Controller Server also acts
as the Data File Server, and the two data storage devices are combined into a central data
storage. However, the preferred embodiment of the present invention is the double-server
configuration depicted in Figure 3. The separation of the Operation System 150 from Render
System 160 allows user interactive operations to continue when the Intelligent Controller
Server 161 needs to be powered down for maintenance or when a system failure occurs to the
Render system 160. Similarly, the maintenance to the Operation System 150 or failure of the
Data File Server 151 does not impact the operation of the Render System 160. In the
configuration depicted in Figure 3, the Central Data Storage 152 stores all image data
required for Pre-processing, Verification and Post-processing. The Controller Data Storage
162 stores all information and data tracked by the Intelligent Controller and acts as a
production database. It also provides a temporary storage for results from render clients
before they are transferred to the Central Data Storage 152.
The Intelligent Controller Server 161 and the Controller Data Storage 162 are the
fundamental hardware devices for the Intelligent Controller 141 that controls all operations in
the Data Processing stage 110. The Intelligent Controller also monitors the process status of
the Data Processing stage and collects information data from other stages that may be
remotely located. Any device or process that accesses data from the Intelligent Controller 141
is treated as a client of the Intelligent Controller Server. This allows the Intelligent Controller
to provide centralized control over the entire re-mastering process, track the status of all
operations from every device and track all image data flow. Through a combination of
software and hardware, the Intelligent Controller 141 performs the following functions:
Resource management - Given a project schedule, it automatically calculates the
minimum daily throughput requirement and manages available resource to meet that
requirement:
• Quality optimization - It automatically determines the optimal set of render
parameters for each render job for the best results. It also automates a quality
evaluation process using statistical measures to determine if render results are
acceptable.
• Computing efficiency optimization - It manages the job queue, schedules and
distributes each job to render clients in the most efficient way with available
computing resources. It provides automated system-wide caching of intermediate data
and process status, based on available storage resource, to minimize time required for
necessary re-render jobs.
• Production management - It tracks and updates all information relating to render
process and flow of image data from every stage of the process and organizes the data
into a database. It produces up-to-date reports on various aspects of the production
process and answers user queries through a query builder.
• System administration - It administrates all render clients and monitors their status,
and it monitors system performance and diagnoses problems.
• User interactivity - It takes user decisions and allows users to overwrite decisions
made by automated schemes. It also makes decisions based on user preference
specified by users.
The details of the four process modules of Data Processing stage are described in the
following sections.
The Pre-processing module 111 is designed as a user-interactive process so that users
can make creative decisions by pre-viewing image data using specially designed software.
The Original Image Data, especially when scanned from film, is typically in the form of long
image sequences with no breaks at scene changes. A major task in the Pre-processing module
is to separate the data into shots, with each shot representing continuity in certain scene
characteristics, such as motion or lighting. Each shot is marked by the start and end frame
numbers and is assigned a unique shot number. In one embodiment of the present invention,
those shots are further grouped into a smaller number of scenes in such a way that all shots
belonging to the same scene share certain common characteristics. Scene grouping makes it
possible to process all shots belonging to the same scene with the same set of parameters. The
shot separation decisions are effectively made by skilled users, but they can also be made
automatically by software through scene analysis.
Another key decision users need to make is re-framing. This operation is generally
needed for a re-mastering project for the following two reasons: first, scanned data typically
includes blank film areas outside the image area that must be cropped for final release;
second, a re-mastered motion picture may be released in a different aspect ratio than it was
originally intended. The re-framing decisions can be effectively made by skilled users. If the
re-framing decisions change from frame to frame within a single shot, a Pan & Scan
operation is needed as part of the re-framing process. Those skilled in the art will readily be
able to perform the operation described by the teaching to satisfy any re-framing decisions.
Figure 4 depicts a typical process of the Pre-processing module. User's decisions can
be made based on direct pre-viewing of Original Image Data, or they can choose to pre-view
a proxy version of the Original Image Data in order to reduce the amount of image data and
shorten data loading time while increasing the run-time for each loading. In one embodiment
of the present invention, the proxy version is created through a proxy generation module 200
that includes downsizing, bit-depth reduction and color correction. The size of the proxy
version must be adequate for users to make Pre-processing decisions while small enough to
ensure high efficiency in software pre-viewing. In another embodiment of the present
invention, a modest data compression is applied to the proxy using standard compression
technology to further increase visual pre-viewing efficiency without affecting viewing
quality. By pre-viewing the proxy version of the image sequence, users can make quick and
key decisions regarding shot editing 201 (the start and end frames for each continuous scene
with a shot number), re-framing 202 (cropping and pan & scan based on a pre-determined
aspect ratio), and color correction parameters 204. Users also select a small number of key
frames 203 from each shot, which will be used in the next Render module for scene analysis.
The Pre-processing also includes an artifact identification process 205 in which artifacts
caused by dirt, film scratch, film degradation and chemical stain, as well as artifacts
introduced in the digital effects stage can be identified and subsequently removed. Those
decisions and the data determined by users are collected by the Intelligent Controller 141
which will apply appropriate image processing software tools to the Original Image Data to
separate them into scene content based shot clips to be ready for render operations. In one
embodiment of the present invention, the image processing tools include shot separation and
file renaming 207, image cropping 208, color correction 209, image resizing 210 and artifact
removal 211.
The proxy version of the Original Image Data is also used for scene classification
206. The purpose of the scene classification process is to group complex scenes into a
relatively small number of classes so that images that belong to the same class share certain
unique characteristics mat makes them different from images of other classes. Scene
classification makes it possible to apply different image processing methods and/or different
render parameters to each class in order to achieve the best results. A sample of scene classes
includes: fast motion scenes, slow motion scenes, still camera shots, dark scenes, bright
scenes, scenes with large portion of sky, face close-up, wide-angle shots, nature scenes, etc.
In one embodiment of the present invention, human users perform scene classification 206. In
another embodiment of the present invention, scene classification is performed by automatic
algorithms based on scene analysis. Each image is represented by a feature vector that may
have a long list of components including color histogram, Grathent histogram, directional
edges, motion, etc. When represented by feature vectors, images are treated as samples in a
multiple dimensional feature space. Standard statistical clustering method can be used to
group samples into preliminary classes. The second pass of motion similarity analysis on the
thumbnail images may ensure samples with continuous motion remain in the same class. A
special set of render parameters can be determined for each class through the render
parameter prediction 220 described in Figure 6.
Artifact removal 211 is necessary for a re-mastering project that demands high image
quality. For Original Image Data scanned from film, artifacts resulting from dirt, dust and
scratches on film are inevitable. Artifacts from film degradation and laboratory chemical
process also occur especially for older film stocks. For Original Image Data available in
original digital form, there may exist artifacts from imperfections in the digital-effects
process. In one embodiment of the present invention, artifacts are removed through an
automated process. Special search algorithms are applied to a range of frames including the
current frame to identify artifacts with known unique characteristics. Pixels identified as
artifacts resulting from dirt and dust display very different intensity values than surrounding
normal pixels and these pixels do not have temporal correspondence from neighboring
frames. Motion estimates of the surrounding normal pixels are computed. As dirt and dust
artifacts are identified, they are replaced by predicted pixel values calculated based on the
motion estimates of the surrounding normal pixels. Pixels identified as artifacts resulting
from film scratches display very different intensity values than surrounding normal pixels,
and these pixels are typically form thin vertical or horizontal lines (depending on film
formats). These film scratch pixels can find strong temporal correspondence from
neighboring frames. As film scratch artifacts are identified, they are replaced by predicted
pixel values calculated by an interpolation algorithm based on surrounding normal pixels.
The automated method is effective when the unique characteristics of the artifacts can be
clearly defined. In another embodiment of the present invention, artifacts are removed by a
semi-automated method. In this method, human users are needed only to identify a small
image region that contains one or more artifacts present in image data using specially
designed software 205, and the locations of those image regions are sent to the Intelligent
Controller 141. Then a searching algorithm is applied to these identified small regions where
artifacts are located to locate pixels with abnormality within the small regions. As artifact
pixels are located, these pixels are replaced by predicted pixel values in the same way as in
the automated embodiment of the present invention that was described. Since the motion
estimation is confined within the identified small regions, the searching and removal
algorithms can be completed within a very short time. For those artifacts that cannot be
removed by either method, they will be re-touched by human users using standard re-
touching software.
In one embodiment of the present invention, artifact removal 210 is performed both in
the Pre-processing module 111 and in the Post-processing module 114, and a majority of
artifacts are removed in the Pre-processing module. In another embodiment of the present
invention, artifact removal is performed only at the Post-processing module. In the latter case,
if a shot is rejected during the Approval stage 130 and needs to be sent back to the Render
module 112 for re-rendering, then the artifact removal operation must be repeated. This will
reduce operational efficiency if manual and semi-automated methods are used.
The enhancement of the image data diat has been separated into shots takes place at
the Render module 112. The operation of the Render module is controlled by the Intelligent
Controller 141 and is totally automated. Figure 5 depicts the process flow diagram of the
Render module. The hardware implementation of the Render module is the Render system
160 diat is equipped with multiples of computing clients 163 configured into clusters (as
shown in Figure 3). Each computing client is a standard computing device. In one
embodiment of the present invention, each computing client is a Pentium processor computer
running on Linux operating system. When an image shot sequence, or a shot, is undergoing a
series of image enhancement operations by render clients, it is referred to as a render job, or a
job. A job can be distributed to a single render client or to a number of render clients. Each
image enhancement operation provides one or more parameters that can be adjusted to
achieve desirable results. The collection of all parameters from every enhancement operation
forms a render parameter set, which determines the performance and the quality of render
results. For image shots with different characteristics, the render parameter set must be
adjusted accordingly in order to achieve the best possible visual quality.
The render parameter set contains parameters crucial for processing on the render
clients 223, and these parameters include those for motion estimation (matching region of
support, the number layers in the hierarchical motion model, searching range, thresholds for
finding the correct matching, etc.), temporal filtering (temporal window size, filter
coefficients, etc.) and sharpening. These parameters can be determined in a number of ways.
In one embodiment of the present invention, the parameter set is predicted by a skillful user
based on visual inspection of every image shot. This approach is only effective if the user's
decisions are consistent and reliable. In another embodiment of the present invention, the
render parameter set is estimated by a prediction algorithm based on image analysis 220.
The algorithm for render parameter prediction is depicted in Figure 6. In order to
reduce computation, image analysis is performed on a sample of image frames selected from
a shot sequence. These sample frames are called key frames, and they are selected by users at
the key frame identification process step 203 in the Pre-processing stage. In another
embodiment of the present invention, key frames can also be determined by an algorithm 300
that computes the histogram of each frame in the shot and determines the most representative
frames based on histogram consistency.
A series of image analysis operations are applied to the selected key frames. In one
embodiment of the present invention, those operations are for the purpose of estimation of
initial motion estimation settings 301, including:
• Estimation of granularity of the image noise distribution 302;
• Estimation of matching region of support (MRS) based on the estimated noise
granularity 303;
• Estimation of global motion by calculating the average absolute motion between key
frames 304; and
• Estimation of searching range based on estimated global motion 305.
Using the estimated matching regions of support and estimated searching range, a
motion estimation algorithm can be applied to key frames to compute the motion estimates
from key frames 306. A variety of motion estimation algorithms are applicable for this
application, and those algorithms are described in the following publication: Christoph Stiller
and Janusz Konrad, "Estimating Motion in Image Sequences", IEEE Signal Processing
Magazine, Vol 16, No.4, July 1999. In one embodiment of the present invention, a pixel-
based motion estimation algorithm based on a hierarchical motion model is implemented.
Based on computed motion vectors at multi-resolution layers, some key statistical
measures regarding image temporal characteristics are calculated 307, and they include:
• Temporal Signal-to-noise Ratio (TSNR) 308 - TSNR measures the level of
temporally uncorrelated noise between key frames. TSNR is computed, after motion
estimation, by warping one of the key frames to the others based on motion vectors,
and then calculating the inverse of MSE (mean square error). TSNR is measured in
dB, similar to conventional SNR used for signal analysis. If TSNR is high, then the
temporal noise is low, and vice versa.
• Motion 309 - The amount of motion that exists in a shot is measured in two ways. In
one embodiment of the present invention, motion is measured by the average absolute
magnitude of motion vectors of all pixels that can be reliably tracked between key
frames. In another embodiment of the present invention, motion is measured based on
the average motion of feature points. Feature points generally provide more reliable
motion estimates than average pixels. However, estimating the motion of feature
points requires a different algorithm and the process must be implemented separately
from pixel-based motion estimation.
• Fast Matching Distribution (FMD) 310 - FMD is a distribution of percentage of direct
pixel match (fast match) between two key frames vs. matching threshold values. The
direct match of pixels between two frames occurs when a matching is found within a
pre-defined small search region. FMD is an indicator of the performance of a motion
estimator. For a given FMD, mere is a set of corresponding threshold values that
delivers the desirable performance.
A prediction algorithm 311 predicts render parameters based on FMD, TSNR and
Motion. In one embodiment of the present invention, the prediction starts with a set of
matching threshold values corresponding to a given FMD. That set of those threshold values
is weighted down when TSNR. is high or when Motion is high, and weighted up when TSNR
is low or when Motion is low.
The operation of the render parameter predictor 220 is controlled by the Intelligent
Controller 141. User input can be fed into the parameter predictor 311 to modify predictions
according to the preference of the users. The predictor 220 also allows users to overwrite its
prediction with a preferred set of render parameters.
Turning back to Figure S, once a render parameter set is determined for an image
shot, it can be submitted to the render queue 221 as a single render job. The Intelligent
Controller 141 checks the validity of the submission and adds the job to a job queue 221. It
then sends the job submission off to one or more available render clients for processing based
a pre-determined job distribution scheme 222. The render clients 223 process the data
according to instructions and compute statistical quality indicator 323. The clients also
frequently report back to the Intelligent Controller 141 their current situation and status. Once
the job is completed, the Intelligent Controller collects distributed results from corresponding
clients and temporarily stores them on the Controller data storage 162. Then it checks the
integrity of the data and the completeness of the job 225. If the Intelligent Controller 141
finds certain frames are missing, or certain frames are incomplete, it requests the render client
to re-process those frames. If it is satisfied that the job is complete, it performs an evaluation
226 to measure the quality of the rendering. If the required quality is not achieved, the
Controller attempts to modify the render parameter set 230 and sends the job back to the job
queue. In special cases, it requests that certain temporal filtering options (see Figure 7 and
Figure 9) be specified 231. For each job re-submission, the version of the job is updated
through a version control scheme 232. If the Intelligent Controller is satisfied with the render
results, it sends the image data to the Central Data Storage 152 and sends a verification notice
227 to the Data File Server 151.
The processing performed by each render client 223 on image shot data consists of a
series of image processing operations as depicted in Figure 7. The major operations include
temporal filtering 322, resizing 326 and sharpening 327. The render client also computes
statistical quality indicators 323 including pixel matching ratio (PMR) and absolute still ratio
(ASR). These quality indicators are used to evaluate render quality in the quality evaluation
stage 226 in Figure 5. In addition, a proxy version of the enhanced image shot is created to be
used for visual inspection at the Verification module 113 in Figure 2. Once a job is
completed, the render client notifies the Intelligent Controller 141, and it will store the results
at the Controller Data Storage 162 and free the render client. The steps in the process
pipeline of a render client are discussed below. All of the steps are not necessarily used in
every application. Figure 7 illustrates the preferred order of the process steps in one
embodiment.
Temporal filtering 322 is the most computationally intensive operation currently
implemented in the render client processing pipeline. It provides two functions: resolution
enhancement 340 and noise reduction 341. The concept of resolution enhancement was
developed based on images originating on photographic film, but the methods developed
based on the concept are not limited to film-based image applications. In fact, the resolution
enhancement method described in the present invention is applicable to any image sequence
captured using some form of optical device, like photo detectors, electronic sensors and
computer-generated images that have been transferred to the above media.
Photographic film contains tiny light sensitive crystals of silver halide salts. When the
film is developed these crystals are turned into tiny filaments of metallic silver. In a color
film, tiny blobs of dye are formed along with the silver during the development process. The
silver is then chemically removed from the film, leaving only this dye image. It is these small
specks of dye that form film grain. Film grains vary in size and shape and are randomly
distributed. But the randomly distributed film grains actually are the fundamental building
blocks that form images on film.
The information contained in a single frame of image can be completely described by
its frequency domain content. When an image frame is digitized, the information retained in
the digital version is limited by the sampling theory. Digitally enlarging a digital image frame
using resizing methods does not add more information, but it makes room in the frequency
domain so that additional image details can be added to the image frame if those details can
be recovered elsewhere. This concept of resolution enhancement is depicted in Figure 1.
Figure 1(A) illustrates the frequency response of a band-limited image signal sampled at
Nyquist frequency Ft > 2F; and Figure 1(B) illustrates the frequency response of the same
image signal with additional high-frequency components recovered from resolution
enhancement process. The image signal is sampled at F',>>2F, and it makes it possible
to add high-frequency details recovered from other image frames. The spatial resolution of
the enhanced image signal now becomes F> F
In one embodiment of the present invention, additional image details are recovered
from neighboring frames through temporal filtering. In temporal filtering, the content of an
image frame is enhanced by a mathematical algorithm that uses information from
neighboring image frames, as depicted in Figure 8. The number of neighboring frames
(including the present frame) used by the temporal filter is referred to as a "temporal
window". For example, the size of the temporal window 401 used by the temporal filter 400
in Figure 8 is 2N +1. By moving the temporal window 401, successive image frames are
enhanced by the temporal filter 400.
The method of resolution enhancement 340 described in the present invention consists
of three major steps: motion estimation, motion field regulation, and detail recovery. A
motion estimation algorithm with sufficient accuracy and robustness is the first essential step.
The existence of random film grains in varying sizes and shapes in photographic images
presents challenges to any motion estimation algorithm. A motion estimation algorithm must
be able to distinguish real image content from random film grain details. A range of well-
known motion estimation algorithms are qualified for this application, and most of them are
described in the following publication: Christoph Stiller and Janusz Konrad, "Estimating
Motion in Image Sequences", IEEE Signal Processing Magazine, Vol 16, No.4, July 1999. In
one embodiment of the present invention, an algorithm based on a hierarchical motion model
is used to achieve both reliability and accuracy. In this method, motion is modeled at multiple
levels of details, making it possible to find major motion properties first and refine the
estimate to details later. In the hierarchical motion model, every image frame is represented
by a multi-level data structure, each representing a specific level of image details. Film grain
details are mostly represented at the lowest level of the hierarchy. Motion estimates
calculated at the top levels of the hierarchy represent real image contents. At each level of the
hierarchy, the motion estimate of every pixel is computed using a variable-size block
matching algorithm for all frames within a temporal window. The searching strategy used in
block matching varies from an exhaustive search to sub-optimal fast search. In an exhaustive
search, all candidates within a predefined searching range will be examined, and the best
match is the one mat minimizes a prediction error. In a fast search, the "best match" criterion
is replaced by a "first match" criterion which takes the first candidate with a prediction error
below a certain set of threshold criteria as the estimate. The "coarse" motion estimate
obtained at the top level is computed as the best match among all candidates. This motion
estimate is successively refined over subsequent levels of hierarchy. This hierarchical search
strategy confines the search within a relatively smaller range at each level so that the
computational complexity will be significantly reduced. At the bottom level, where film grain
structure becomes significant, it is necessary to apply a search strategy that is robust to film
grain variance. Since film grain size changes according to film stock, lighting conditions,
laboratory process, etc. the matching regions of support must be sufficiently larger than the
maximal film grain size.
Motion field regulation is necessary since motion estimation is an ill-posed problem
and multiple solutions exist given a set of searching criteria. Most estimation errors occur in
smooth regions of images where the size of potential solution space increases drastically. The
approach is to constrain the solution space using such constraints as high-frequency features,
smoothness and quality measure. High-frequency features are "visually significant" image
features that can be identified through feature analysis, and they represent significant changes
in pixel intensity, pixel color or pixel morion. The motion estimates obtained from pixels
representing high-frequency features are more reliable and more accurate than those obtained
from pixels in smooth region. The smoothness constraint states that neighboring pixels in a
smooth region are most likely to share similar motion estimates. For high-frequency features,
the smoothness should apply in the direction of feature orientation. Each motion estimate is
assigned a reliability measure indicating the reliability of its estimate. The reliability measure
of a motion estimate is in inverse proportion to the size of solution space associated with the
estimate. For a pixel with a lower reliability measure value, motion field regulation should be
considered, and more constraints should be applied to its solution space in order to reduce the
estimation error.
To achieve sub-pixel accuracy at a given frame interval, a group of synthesized
frames are constructed by mapping each neighboring frame to the present frame intervals
based on the corresponding motion estimates. An error map between each synthesized frame
and the present frame is calculated. For those pixels with large estimation errors, their motion
estimates are modified until the errors are minimized. Further improvement of estimation
accuracy can be achieved using more sophisticated motion models like six-parameter affine
models or eight-parameter projective linear models. The synthesized frames are re-created
based on the modified motion estimates, and they will be used in the detail recovery step.
A resolution-enhanced image 403 is constructed through adaptive temporal
interpolation of synthesized frames within the temporal window 401. An adaptive temporal
filter 400 with FIR (finite duration impulse response) structure is applied to each pixel
location. The filter coefficients are determined based on the motion estimate reliability
measure associated with each pixel. If a motion estimate is highly reliable, the filter
coefficients are weighted in proportion to the reliability measurements from each frame
interval. On the other hand, if a motion estimate is unreliable, the filter coefficients are
weighted more heavily on the current frame.
Although the previous description of the temporal filtering is based on photographic
images, it equally applies to images captured using some form of optical device including
photo detectors and electronic sensors as well as computer-generated images that have been
transferred to the above media.
A direct result from resolution enhancement 340 is the reduction of the visual
sensation of "boiling" film grain in the enhanced image sequence 403. As noted earlier, film
grain varies in size and shape and is randomly distributed spatially and temporally. This
random pattern creates the "boiling" grain phenomenon when images are viewed in motion.
In a large format cinema, audiences are typically seated closer to the screen than they are in a
regular cinema so that images cover a much wider portion of their field of view. This results
in the "boiling" film grain becoming visibly objectionable and is perceived as undesirable
noise. The temporal filtering method 400 suppresses "boiling" film grain noise because it
improves temporal correlation between neighboring image frames in a sequence.
Furthermore, film grain is also reduced as the result of enhanced spatial resolution.
The single-pass temporal filtering algorithm depicted in Figure 8 can be implemented
in parallel processing mode since each output frame can be computed independently. For a
majority of image shots, the single-pass temporal filtering is very effective in noise reduction.
However, for very noisy images, or for jobs where noise level is a major concern, other noise
reduction algorithms 341 can be applied. In one embodiment of the present invention, noise
reduction 341 is achieved with multi-pass temporal filtering following one of the three
methods depicted in Figure 9. The actual temporal filter device 400 is omitted in Figure 9 for
clarity of the description, as one skilled in the art will readily be able to implement those
algorithms described by the teaching.
Figure 9(A) describes a multi-pass temporal filtering algorithm 410. The multiple-
pass algorithm basically repeats the single-pass scheme multiple times. Render parameters
prediction described in Figure 6 is applied before each pass. The multi-pass temporal filtering
is effective in suppressing noise in very noisy images.
Figure 9(B) describes a multi-pass algorithm 411 based on a temporal pyramid. The
"temporal" sampling is "coarse" in the first pass and every N number of frames within a
temporal window is used. The temporal sampling rate increases for the next pass until every
neighboring frame within the temporal window is used in the final pass. Temporal pyramid
filtering tends to reduce the computation required for the scheme described in Figure 9(A).
Figure 9(C) describes a serial temporal filtering algorithm 412 that can be
implemented as an iterated process. The previously processed frames are used immediately
for processing the next frames. The algorithms depicted in Figure 9 generally perform better
in noise reduction, but they tend to reduce image details compared with the single-pass
algorithm of Figure 8. Certain criteria can be defined in the temporal filtering options process
231 (as shown in Figure 5) to select the most appropriate options for temporal filtering based
on image noise level represented by the TSNR that is calculated at process 308
Figure 5 shows mat the render quality evaluation process 226 is implemented to
evaluate render performance based on statistical quality indicators computed by render clients
during temporal filtering 323. In one embodiment of the present invention, the quality
evaluation algorithm is based on PMR (pixel matching ratio) and ASR(absolute still ratio).
PMR and ASR measures the impact of actual temporal filtering operations performed to each
image frame, and they are good indicators for render parameter predictions that are out of
normal range. The PMR and ASR are used to predict render parameter modifications required
to achieve better results.
PMR measures the percentage of pixels that were reliably estimated over a range of
frames within the temporal window. In temporal filtering, not every pixel can find
correspondence over every frame inside a temporal window. Some pixels can only be tracked
over fewer frames, and some pixels are unable to be tracked at all. PMR is a distribution of
the percentage of each type of pixels over the range of frames that they can be tracked. A
PMR that peaks at the high end (more frames) indicates that the majority of pixels are
reliably tracked over the entire temporal window. On the other hand, a PMR that peaks at the
low end (fewer frames) indicates that the motion estimation algorithm has trouble in tracking
that image. A high PMR distribution suggests that the current image frame is relatively clean,
but it can also indicate an improper render parameter setting (matching thresholds too higher,
MRS too small, for example). A low PMR distribution may suggest a very noisy image, but it
can also indicate incorrect parameter settings (low matching thresholds, MRS too large, for
example).
ASR measures the distribution of pixels that can easily be tracked over a range of
frames without significant searching effort. Those pixels that remains at the same location
throughout the entire range of temporal window are called "absolute still pixels". In another
embodiment of the present invention, the absolute still pixels include those with small
changes in their positions. The peak of ASR is high for images with little motion, and it is
low for images with significant amount of motion. However, if ASR is high for images with
reasonable amount of motion, then it may indicate improper setting of render parameters
(thresholds too high, search range too small, etc.).
In one embodiment of the present invention, the render quality evaluation process 226
is an automated process. For each class of scenes determined at the scene classification
process 206, a standard profile of PMR distribution is pre-determined which represents
average PMR distributions of scenes in the same class. Similarly, a standard profile of ASR is
also determined. The evaluation algorithm 226 compares the PMR and ASR of the current
job with those standard profiles and determines if temporal rendering applied to the job is
within a normal range. If the current job has a very different PMR or ASR distribution
compared with standard profiles, then the job is rejected and sent back to the same render
clients for re-rendering with a modified set offender parameters.
The render parameters estimated by the parameter predictor 220 might not be optimal
for local variations of image characteristics. One example is noise distribution. It is well
known that film grain has a non-uniform distribution vs. film density. Film grain is most
evident in the midtones of a print, which presents a density range about 0.6 to 0.9. Film grain
also tends to increase when camera exposure decreases. Similar non-uniformity exists for
images captured using an electronic camera, where noise level tends to be high in darker
areas. The non-uniformity of the noise distribution can also be the result of non-linear
conversion (logarithmic, gamma, etc.) commonly used in file format conversion.
Turning again to Figure 7, in order to effectively suppress noise in darker areas
without affecting image quality in other regions, a noise equalization algorithm 321 is applied
to images to keep noise distribution uniform before temporal filtering. The algorithm consists
of the following steps. The first step is to calculate the noise histogram. A full-search motion
estimation is performed on a pair of frames selected from key frames using a set of relatively
large thresholds. Then block matching errors are calculated for every pixel. For those pixels
with a low block match error, calculate the histogram and compare it with a standard
histogram. The noise histogram is a good representation of the noise and pixel luminance
relation. The MSE between the noise histogram and a uniformly distributed histogram is
calculated. If the MSE value is large, then the noise distribution of the images are considered
non-uniform. Find a transform (in a form of a lookup table) that can equalize the noise
histogram, and ensure it also has an inverse transform. Finally, apply the transform to the
images before temporal filtering. After the temporal filtering, the inverse transform 324 must
be applied to the images in order to retain its original color.
The underlining assumption of a motion estimation algorithm is that there is no abrupt
color change between neighboring frames. However, this assumption breaks down when
there is lighting change in a scene, such as lightning, flickering campfire, moving shadows,
etc. Without a proper compensation, the effectiveness of temporal filtering 322 will be
reduced. In one embodiment of the present invention, the lighting compensation algorithm
320 starts with selecting the brightest frame as the reference, and then it tries to find a
transform for every frame to match the reference. Those types of transforms include gamma
curves, histogram stretch or other monotonic non-linear transform. The transforms can be
found by matching histograms of a frame to that of the brightest frame. In order to maintain
smooth transitions between frames, a temporal filter can be applied to the transforms to
ensure temporal consistency. Once the transform for each frame is found, apply it to each
frame before temporal filtering, then apply the inverse transform 325 to the enhanced data to
retain their original color. This method is suitable for sequences with frequent but relatively
small lighting changes.
In another embodiment -of the present invention, lighting change compensation can
also be implemented as an adaptive mechanism of temporal filtering 322. In this method, the
histogram of every frame is compared with every neighboring frame within the temporal
window. Then all neighboring frames are compensated against the current frame as the
reference in a similar method, as described in the previous paragraph. Since the algorithm
always uses the current frame as the reference, no inverse transform is needed for this
adaptive lighting compensation algorithm. This method is suitable for scenes that contain
infrequent but strong light changes.
If noise equalization 321 and lighting change compensation 320 are both needed for a
render job, the lighting change compensation 320 should be applied before the noise
equalization 321. In another embodiment of the present invention, the transforms of both
stages are combined into a single transform so that it can be applied only once.
Sharpening 327 is the last step of the render client pipeline, and it emphasizes the high
frequency components of an image. Since the recovered image details from resolution
enhancement are mostly high frequency components, sharpening can significantly improve
image quality. Sharpening can also be modeled as a process to recover MTF loss over the
process of image formation. MTF represents modulation transfer function, and it is used to
analyze the quality of an imaging system. For example, the quality of images resulting on
film can be modeled as multiplication of the MTF of individual devices in the process. These
devices may include camera optics, negative film, printer film, and printer and scanner optics.
Since the majority of these devices has low-pass MTF, the combined MTF of the imaging
formation process must have a low-pass MTF. Therefore, a desirable sharpening algorithm
should display high-pass MTF characteristics in order to correct the degradation of image
quality.
A standard unsharp mask filter is such a sharpening algorithm, and it can be described
by
where LP(f(x,y))is a low-pass filter. The tennf(x,y)-lP(f(x,y)) exhibits a high-pass
characteristic, and the unsharp mask filter boosts it by multiplying with a sharpen
gain g(x,y). The boosted high-frequency components are added to the original image f(x,y).
The filter gain g(x,y) is usually a constant, but it can be made adaptive based on local
characteristics. At smooth regions, small filter gain should be selected so that the unwanted
characteristics, like film grains, will not be emphasized.
One problem with the standard unsharp mask filter is that the range of high-frequency
components that are emphasized is limited by the kernel size of the low-pass filter in equation
(1). By varying the kernel size of the low-pass filter, different levels of image details,
corresponding to different sections in the MTF curve, can be selectively emphasized. For
motion picture images that usually contain a relatively large range of image details, it is
important that sharpening improves system MTF over a relatively broad range of detail
levels. To achieve that goal, the present invention generalizes the unsharp mask filter of
equation (1) to support multiple levels of details by the following description:
In equation (2), the sharpening gain value gt for the itth detail level can be selected to
compensate for the MTF degradation at that particular detail level. In one embodiment of the
present invention, Gaussian low-pass filters are used in equation (2), and up to six levels of
kernel sizes are deployed. Those skilled in the art will recognize that the present invention is
not limited to Gaussian filter and six levels of details. Other types of low-pass filters and
more detail levels are possible following the teaching of the present invention.
The Render module depicted in Figure 5 is specially designed for achieving high
efficiency for image processing tasks that require temporal operations. The Intelligent
Controller 141 manages render job distribution and assigns jobs to specific render clients
based on a pre-determined load-balancing scheme. If there are multiple available candidates,
the Intelligent Controller checks the network traffic load distribution among render client
clusters 163 and selects a render client (or render clients) from the cluster (or clusters) with
the lowest traffic load. For each job in a queue, it may assign it to a single render client,
especially when mere are more jobs waiting in the queue than the number of available render
clients, or it may assign the job to a multiple of render clients, especially when the job needs
to be completed as quickly as possible. In one embodiment of the present invention, the job
distribution process 222 follows one of the three schemes depicted in Figure 10.
In Scheme A 420, each shot is always assigned to a render client, and it will be sent to
a render client with the shortest waiting time. For instance, if two render clients are available,
the job will be assigned to the render client which has fewer frames in waiting to be
processed. If both render clients have the same amount of waiting time, the job will be sent to
the render client whose cluster has the least amount of load. In Scheme B 421, a single shot
is split into a number of segments, and each segment contains at least a minimal number of
frames. Each segment is distributed to a render client following the same "shortest waiting
time and least amount of load" criteria. There should be sufficient frame overlapping between
segments so that each segment is correctly rendered by temporal filtering. The amount of
overlapping frames needed is determined by the temporal window size. The Intelligent
Controller 141 must always be aware of the current temporal window size used in the
temporal filtering and calculates the required overlapping frames. In Scheme C 422, each
frame is further divided into regions, and each region is distributed to a render client. Due to
the nature of motion estimation, sufficient overlapping rows and columns must be allowed for
in each region in order to accommodate the search strategy deployed by the motion
estimation algorithm. Those skilled in the art will recognize that the present invention is not
limited to the three schemes depicted in Figure 10, and other job distribution schemes are
possible following the teaching of the present invention.
Each render client, once instructed to run a job, is responsible for pulling all image
data it requires from the Central Data Storage 152, executing required operations on each
frame and pushing the enhanced image data to a temporary location at Controller Data
Storage. For a job that was distributed to multiple render clients, the Intelligent Controller is
responsible for assembling 224 rendered segments from render clients into a continuous shot
The Intelligent Controller also checks the integrity 225 of the assembled data for occasional
missing frames or incomplete frames in the shot If missing frames or incomplete frames are
discovered, the Intelligent Controller sends a request to the same render clients for re-
rendering of those frames. The communication between the Intelligent Controller and render
clients is crucial for render efficiency. The Intelligent Controller tracks the current state of
each render client and constantly monitors for available processors. In the eventuality of
failure of a render client, the Intelligent Controller raises an alert for repair. It reroutes the job
to other available clients for processing. A Diagnostics process ensures that there is no loss of
data during the transfer. If the Intelligent Controller server experiences a failure, the state of
the system before malfunction is preserved. In one embodiment of the present invention, the
Intelligent Controller server re-starts by killing all processes that are running on render clients
and re-assigns jobs to each render client. In another embodiment of the present invention, the
Intelligent Controller polls the render clients for their status, finds their current states and
resumes the control. This is a more complicated re-start scheme, but no re-rendering of data
is required.
As described in earlier sections, the performance of the render operations is evaluated
by evaluating statistical quality indicators, like PMR and ASR, calculated by the render
clients. However, a normal PMR or ASR measure does not guarantee the optimal visual
quality. Human visual inspection is needed to ensure final visual quality, and this process is
implemented in the Verification module 113, as depicted in Figure 11. A proxy version of the
enhanced image data, which was generated in the proxy generation process 328 in the render
client pipeline (Figure 7) is used for visual inspection. The size of the proxy images should
be adequate for users to spot existing render problems while small enough to ensure software
viewing efficiency. By viewing the proxy version displayed in real-time with special purpose
software, users are able to make decisions on image quality.
In one embodiment of the present invention, users first check if each job is complete
240. The automated integrity check process 225 at the Render module does not capture all
render problems, and those frames missed by the integrity check 225 will be caught at this
stage. Those frames found with problems are re-submitted to the original render client for
processing. Once a job is deemed complete, users will check the following quality aspects
that are of the most concern for users:
• verify if the noise level of overall image and in local regions are acceptable 244;
• verify if visual sharpness are appropriate 245;
• verify if sufficient image details are preserved and enhanced 246;
• verify if the re-framing decisions are correct 247;
• verify if further color correction is needed 248;
• verify if there exist artifacts that need to be removed 249;
• verify if the motion exists in the shot may cause viewing discomfort 250.
The inspection of image details requires users to view images at its full resolution. In
that case, both the enhanced image data and Original reference are available to users. In one
embodiment of the present invention, special purpose software displays both image data
within the same viewing window so that users can compare two images using digital wipe
function.
If users find that the noise level of a job is too high, or visual sharpness of the job is
not appropriate, or there is unacceptable loss of image details, the job will be re-submitted to
the Render module with modified render parameters 251. In one embodiment of the present
invention, the decision is made by users consulting with measured statistical quality
indicators. The Intelligent Controller 141 provides users with a graphical display of those
statistical measures (PMR, ASR, TSNR, Motion, etc.) on their workstations 153 through
special-purpose software. Users make educated decisions about necessary parameter
modifications based on the available statistical data. In another embodiment of the present
invention, an automated algorithm is implemented in the process 251 to compute the
necessary modifications to render parameters based on the same set of statistical measures.
If users find problems with re-framing, the image shot will be sent back to the Pre-
processing module to obtain new re-framing decisions. For problems concerning color 248,
artifacts 249 and motion correction 250, users make correction decisions 252 and send the
image data to Post-processing for fixing without a pass stamp 253. Every image shot with
acceptable image quality gets a pass stamp 254 from the Intelligent Controller and is also sent
to the next Post-processing module for data output.
As shown in Figure 12, at the Post-processing module, image shots that require fixing
are sent for artifact removal 265, or for color correction 266, or for motion correction 267.
The methods for artifact removal and color correction are very similar to processes at the Pre-
processing module 111. The fixed shots are sent back to verification 113 to get pass stamps
254. Any image shot without a pass stamp is not allowed to pass checking point 260. The
approved image shots are organized in the same order 261 as in the motion picture and
converted to the required output format before being sent to the Image Out stage 120. All
these operations at the Post-processing stages are controlled and tracked by the Intelligent
Controller 141.
Motion correction 267 is a process specific to large format projection requirement.
When a conventional motion picture is exhibited in a large format cinema where images
cover a much larger portion of an audience's field of view, the sensation of motion in the
motion picture is also magnified. For scenes that contain fast camera motion or rigorous
object motion, the magnified motion sensation may cause viewing discomfort for some
audiences. Motion correction is a method to reduce motion-related viewing discomfort by
reducing angular movement.
In one embodiment of the present invention, the motion correction method is to
reduce two types of motion problems: motion strobing and extreme camera shaking. Motion
strobing is the perceived motion discontinuity caused by a fixed projection frame rate. The
method of reducing motion strobing is to add motion blur to images without increasing the
projection frame rate. Motion blur can be generated by applying a directional low-pass filter
in the direction of motion for moving pixels. The direction of motion can be retrieved from
motion estimates already calculated in the temporal filtering process 322.
Extreme camera shaking can be reduced by partial camera stabilization. The motion
of the camera can be calculated by tracking multiple feature points located in the image
background. Starting from a large number of features, several thousands for example, the
tracking algorithm eliminates most of those features until only the most reliable features are
left Then the process repeats in subsequent frames until the end of the shot. In this way, the
most common features are found throughout the entire sequence. For each feature, a motion
vector between adjacent frames may be defined. A statistical clustering method is used to
group features into regular moving features and irregular moving features. The global camera
motion curve is then calculated by averaging all regular moving features. Camera
stabilization is achieved by reducing the global motion curve and calculating the entire scene
based on tracked features. The amount of motion reduction is the result of the tradeoff
between reducing viewing discomfort and maintaining the same motion sensation that
filmmakers originally intended.
One important aspect of the present invention is mat the Intelligent Controller 141
provides a function of production management, which is extremely important for the success
of a motion picture re-mastering project. Since every device and process that accesses data
from the Intelligent Controller 141 is treated as a client, the client-server configuration allows
the Intelligent Controller to manage the progress of the entire re-mastering project and to
track the status of every operation in every stage of the process.
In one embodiment of the present invention, the types of information that are tracked
by the Intelligent Controller 141 are listed in Figure 13. Examples of production data that are
tracked include:
• Status of Original Image Data;
• A scene/shot list and future modifications;
• Pre-processing decisions for each shot;
• Status of every shot in the different stages of the process;
• Operations applied to each shot in every stage of the process;
• Parameters used for render operations for each shot;
• Version status of each shot;
• Verification decisions of each shot;
• Status of image out process for each shot;
• User preference and user decisions related to each shot;
• User notes related to the processing of each shot; and
• Approval decisions for each shot, etc.
Based on the above information, the Intelligent Controller provides up-to-date reports
regarding the status of the production. The formats of the reports are specified by users.
Examples of the reports include:
• Percentage of film scanned and received;
• How many shots have been processed;
• Project completion date estimation based on current throughput;
• Percentage of shots have been approved;
• List of shots that were rendered with a particular set of parameters;
• How many versions are there for each shot;
• Daily, weekly, monthly throughput report; and
• System utilization report, etc.
The Intelligent Controller also allows users to construct their own reports through a
query system. Examples of the information that the query system supports include:
• Shot numbers;
• Shot length;
• Shot versions;
• Render parameters;
• Verification status; and
• Approval status, etc.
The foregoing is provided for purposes of illustrating, explaining, and describing
embodiments of the present invention. Further modifications and adaptations to these
embodiments will be apparent to those skilled in the art and may be made without departing
from the scope or spirit of the invention.
WE CLAIM :
1. A method for digitally enhancing image resolution and quality of image sequence
data, comprising:
converting the image sequence to a digital format to produce converted image
data that comprises original image data;
processing the converted image data to produce enhanced image data; and
outputting the enhanced image data;
wherein converting the image sequence, processing the converted image data,
and outputting the enhanced image data are automatically controlled and allow user
input and interaction.
2. The method ef as claimed in claim 1, additionally comprising converting the
enhanced image data to an alternative format.
3. The method as claimed in claim 1, additionally comprising approving the output
enhanced image data.
4. The method as claimed in claim 1 wherein the image sequence is a motion
picture originated in a 35 mm film format, and wherein converting the image sequence
to a digital format comprises digitizing the film using a film scanner.
5. The method as claimed in claim 1 wherein the image sequence is a motion
picture originated in a digital format, and wherein converting the image sequence to a
digital format comprises data conversion from the original digital format.
6. The method as claimed in claim 1 wherein the image sequence is in any image
sequence format containing images captured using a device with an optical or
electronic sensor.
7. The method as claimed in claim 2 wherein the alternative format is a film format
with enhanced resolution and quality, and wherein converting the enhanced image data
to an alternative format comprises film recording using a film recorder.
8. The method as claimed in claim 2 wherein the alternative format is a digital
display format with enhanced resolution and quality, and wherein converting the
enhanced image data to an alternative format comprises data conversion to digital
display format.
9. The method as claimed in claim 2 wherein the alternative format is a concurrent
release of an original motion picture with enhanced resolution and quality.
10. The method as claimed in claim 1 wherein processing the converted image data
comprises:
a pre-processing process;
a render process;
a verification process; and
a post-processing process.
11. The method as claimed in claim 10 wherein the pre-processing process
comprises at least one user decision based on previewing the converted image data
and previewing a proxy version of the converted image data and wherein the pre-
processing process comprises:
shot separation;
cropping;
color correction;
image resizing; and
artifact removal.
12. The method as claimed in claim 11 wherein the scene classification is performed
by an automated algorithm based on scene analysis.
13. The method as claimed in claim 11 wherein the artifact removal is performed by
an automated algorithm in which artifacts are identified by their unique characteristics
both from a current frame and from neighboring frames and removed by predicted pixel
values calculated based on inter-frame motion estimates or intra-frame interpolation.
14. The method as claimed in claim 11 wherein the artifact removal is performed by
a semi-automated algorithm for interactively locating artifacts and automated removal
using temporal processing.
15. The method as claimed in claim 10 wherein render process comprises:
receiving converted image data separated into shots having image shot data,
wherein each shot is a render job;
predicting a render parameter set for each render job;
distributing render jobs among multiple computing render clients;
processing the render jobs to produce render results;
assembling the render results into a continuous shot;
checking the integrity of the assembled render results and repairing by re-
rendering missing or incomplete frames;
evaluating processing render quality based on statistical quality indicators of the
assembled render results; and
modifying the render parameters for improving the render results and re-
processing as determined by evaluation.
16. The method as claimed in claim 15 wherein predicting the render parameter set
for a shot comprises:
locating key frames;
computing initial motion estimation settings;
computing motion estimates by using the initial motion estimation settings and
applying motion estimation algorithms to the key frames; and
computing statistical temporal measures based on the computed motion
estimates to generate the render parameter set; and
determining render parameters using statistical temporal measures and user
preferences.
17. The method as claimed in claim 16 wherein computing statistical temporal
measures comprises:
temporal signal-to-noise ratio;
motion; and
fast matching distribution.
18. The method as claimed in claim 41, wherein the temporal filtering schemes
comprises:
computing motion estimates;
regulating motion fields; and
recovering details.
19. The method as claimed in claim 15 wherein the statistical quality indicators
comprise pixel matching ratio and absolute still ratio.
20. The method as claimed in claim 15 wherein the evaluation of processing render
quality is based on pixel matching ratio and absolute still ratio.
21. The method as claimed in claim 15 wherein the render parameters are modified
based on pixel matching ratio and absolute still ratio.
22. The method as claimed in claim 41, wherein the modulation transfer function
compensation is based on selective enhancement of multiple levels of image details.
23. The method as claimed in claim 15, wherein distributing each render jobs among
multiple render clients is performed based on a load-balancing scheme by at least one
of:
assigning an entire image shot to a single render client;
dividing an image shot into overlapping segments and distributing them to render
clients; or
dividing an image frame into overlapping regions and distributing them to render
clients.
24. The method as claimed in claim 10, wherein the verification process comprises:
visually checking a proxy version of the enhanced image data for completeness;
visually verifying image noise level, sharpness and detail preservation and
modifying render parameters for re-submission if necessary based on statistical
measures;
visually verifying color, artifacts and motion comfort level; and
issuing a pass stamp if a render image shot meets quality standard.
25. The method as claimed in claim 10 wherein the post-processing process
comprises at least one of:
final artifact removal;
final color correction;
motion correction;
organizing images; or
output image data conversion.
26. The method as claimed in claim 25 wherein motion correction comprises:
reducing motion strobing by adding motion blur; and
reducing extreme camera shaking by partial camera stabilization.
27. A system for digitally enhancing resolution and quality of an image sequence,
comprising:
an intelligent controller that automatically controls, monitors, and manages
digitally enhancing the resolution and quality of an image sequence, the intelligent
controller comprising:
a central control computer server; and
a controller data storage device for supporting the central control computer
server;
a data file server configured as a client device of the central control server;
multiple computing render devices for processing original image data to produce
enhanced image data having enhanced resolution and quality, the render devices
being automatically controlled and adapted to receive user input and interaction and
configured as client devices to the central control server;
multiple workstations for user input and interactive operations relating to the
processing of the original image data and configured as client devices to the central
control server;
a central data storage device for storing image data and process data;
image data input and output devices; and
a computer network that provides communications between the central control
server and all client devices.
28. The system as claimed in claim 27 where the computing render devices are
standard computers configured in a parallel and distributed configuration designed for
supporting temporal processing of multiple image sequences simultaneously.
29. The system as claimed in claim 27 wherein the central control computer server
is capable of resource management; quality optimization; computing efficiency
optimization; production management; system administration; and user interactivity.
30. A temporal image processing method for enhancing the image resolution and
quality of an image sequence, comprising:
recovering image tails from multiple image frames using a temporal filtering
method with render parameter prediction;
evaluating the performance of the temporal filtering and modifying render
parameters using statistical quality indicators; and
sharpening image details by emphasizing a selected range of detail levels.
31. The method as claimed in claim 30, wherein the temporal filtering method
comprises the steps of:
computing motion estimates;
regulating motion filed; and
recovering details.
32. The method as claimed in claim 31, wherein the temporal filtering method
additionally comprises at least one of:
a multi-pass temporal filtering method;
a temporal pyramid filtering method;
a serial temporal filtering method; or
a single pass temporal filtering method;
33. The method as claimed in claim 30, wherein the render parameter prediction is
computed based on statistical measures.
34. The method as claimed in claim 33, wherein the statistical measures comprise:
temporal signal-to-noise ratio;
motion; and
fast matching distribution.
35. The method as claimed in claim 30, wherein the temporal filtering method
additionally comprises:
a noise equalization method; and
a lighting change compensation method.
36. The method as claimed in claim 30, wherein sharpening image details
comprises:
using varying kernel sizes to selectively emphasize image details at different
detail levels; and
using a selective sharpening gain value at each detail level to compensate
modulation transfer function degradation.
37. The method as claimed in claim 30, wherein evaluating the performance of the
temporal filtering is based on statistical quality indicators, the statistical quality
indicators comprising:
a pixel matching ratio; and
absolute still ratio.
38. The method as claimed in claim 30, wherein sharpening image details is based,
at least in part, on modulation transfer function compensation that is based on selective
enhancement of multiple levels of image details.
39. The method as claimed in claim 16, wherein computing initial m5tion estimation
settings comprises:
estimation of granularity of the image noise distribution;
estimation of matching region of support based on the estimated noise
granularity;
estimation of global motion by calculating the average absolute motion between
key frames; and
estimation of searching range based on estimated global motion.
40. The method as claimed in claim 15, wherein processing of each of the
distributed render jobs comprises:
one or more temporal filtering schemes additionally comprising at least one of:
using a multi-pass temporal filtering method;
a temporal pyramid filtering method;
a serial temporal filtering method; or
a single pass temporal filtering method;
image resizing; and
image sharpening based on modulation transfer function compensation.
41. The method as claimed in claim 40, wherein processing of each of the
distributed render jobs additionally comprises at least one of:
lighting change compensation and lighting change inverse transform;
noise equalization and inverse noise equalization; or
proxy generation.
42. The system as claimed in claim 27, additionally comprising a central data
storage comprising the controller data storage device and the central data storage
device; and
wherein the central control server comprises the data file server.

A process and methods of digital enhancement of motion pictures and other
moving image sequences for the purpose of being exhibited in an alternative display
format including a large format cinema are disclosed. The invention efficiently
enhances image resolution and quality through a temporal filtering process 322 and
achieves high performance using automated or interactive statistical quality evaluation
methods 323. A system specially designed for efficient temporal computing with a
parallel and distributed computing configuration 150, 160 equipped with a variety of
optimization schemes is also disclosed. The performance of the process and the
system is optimized through an intelligent controller 161 and is scalable to support any
throughput requirements demanded for concurrent motion picture releases in the
original format as well as in any alternative format.