BACKGROUND
This invention relates monitoring audio-visual content with captions.
5 SUMMARY
In one aspect, the present invention consists in a method of monitoring audiovisual
content which includes a succession of video images and a plurality of
caption events, each caption event being associated with and intended to be
co-timed with a respective string of successive images, the method
10 comprising the steps of processing a caption event to derive a caption event
fingerprint; searching audio-visual content to identify a caption event matching
a defined caption event fingerprint; analysing any matching caption event;
and measuring any caption event error.
In another aspect, the present invention consists in a system for monitoring
15 audio-visual content which includes a succession of video images and a
plurality of caption events, each caption event being associated with and
intended to be co-timed with a respective string of successive images, the
system comprising: at least first and second fingerprint generators operating
in a content delivery chain at respective locations upstream and downstream
20 of defined content manipulation process or processes, each fingerprint
generator serving to process a caption event to derive a caption event
fingerprint; and a fingerprint processor serving to compare caption event
fingerprints from the respective fingerprint generators to identify matching
caption events; and to measure any caption event error.
25 The measured caption event error may be selected from the group consisting
of a missing caption event; a caption event timing error and a caption
discrepancy. Timing may be determined relative to the succession of video
images of the identified caption event.
A caption event may comprise a plurality of words, each formed from one or
30 more characters, and the caption event fingerprint may be derived from a
length of each word in the caption event, without regard to the identity of the
2
character or characters forming the word. Where the caption event
comprises a caption image, the length of each word in the caption event may
be determined by: analysing the caption image to identify caption image
regions corresponding respectively with words in the caption; and determining
5 a horizontal dimension of each such caption image region. A caption image
may be analysed to identify caption image regions corresponding respectively
with lines of words in the caption and the length of a word is represented as a
proportion of the length of a line.
The length of a word is represented as a proportion of the length of a line
10 containing the word. Alternatively, a measurement window of audio-visual
content is defined containing a plurality of caption events and the length of a
word is represented as a proportion of the representative line length derived
from the measurement window. The representative line length may be the
average line length in the measurement window or the length of a
15 representative line in the measurement window, for example the longest line,
the line with the greatest number of words, or the temporally closest line.
In the preferred arrangements, the text of a caption event cannot be derived
from the caption event fingerprint.
In some arrangements, a plurality of measured caption event errors are
20 combined to generate a flag indicating whether or not captions are
acceptable.
BRIEF DESCRIPTION OF THE DRAWINGS
The invention will be described by way of example with reference to the
25 accompanying drawings, in which:
Figure 1 is a diagram illustrating captions monitoring across part of a
broadcast chain.
Figure 2 is a diagram illustrating searching to identify a matching caption.
Figure 3 gives an example of a caption at different stages in a broadcast train.
3
Figure 4 is a diagram illustrating caption image analysis.
Figure 5 illustrates caption matching between two text-based caption
fingerprints.
Figure 6 illustrates caption matching between text-based and image-based
5 caption fingerprints.
DESCRIPTION OF PREFERRED EMBODIMENTS
Embodiments of this invention provide for the matching of video captions
originating from different points in a broadcast chain for the purposes of
10 checking their integrity, absolute delay, and delay changes relative to video,
and for the measuring of errors.
Unlike video and audio which both involve fixed sample rates (albeit in
various standards), captions are intrinsically non-periodic. A caption will
generally be associated with a string of video images, with the strings varying
15 in length from caption to caption. This means that the task of correlating two
sources of caption data is fundamentally different to the task of correlating
video and or audio.
Caption errors can take various forms. A caption may be missing; there may
be a timing error or other caption discrepancies. There may be a number of
20 qualitative differences that individually or in combination degrade the ‘user
experience’ by. Missing caption events, character errors, differences in white
spaces, colour, position, or display time differences are all of relevant, but the
importance of each may be subjective. It would be very useful to combine
these into a score which reflects whether a caption channel is being delivered
25 with acceptable quality. This disclosure provides basic underlying
measurements that are necessary to contribute to such a qualitative score,
and makes some suggestions regarding a more general subjective
measurement.
A further issue is that, captions can exist in multiple different formats - some
30 of which are not easy to compare. Therefore a matching approach based on a
4
particular transport protocol or wrapper is undesirable. A considerable
number of caption formats exist, but broadly they occur in two categories;
they can be text-based or image-based. Comparing these two types is a
challenge: for example verifying that the end of a Freeview broadcast
contains DVB captions 5 that are timed correctly against the original text
specification they were derived from. In principle, this type of comparison can
be tackled by first using Optical Character Recognition (OCR) to extract text
from images before addressing the text matching problem. Established OCR
algorithms exist which could be used. But there is naturally a processing
10 overhead – analysis to extract characters and correct errors against a known
dictionary, involves considerable effort. (It does of course provide a route to
dictionary look-up and language detection and more advanced analysis, but
comes at a price).
It would in many cases be advantageous to have solutions to text comparison
15 and image/text matching which removes the need for OCR.
As mentioned, a wide range of caption formats exist. These differ between
delivery systems (traditional broadcast, internet, DVD, Blu-ray, cinema etc)
and also differ between territories. For example, UK Freeview broadcasts
carry image-based DVB captions, whereas US captions are broadcast in text20
based EIA-608 or CEA-708 format, and in Europe the text-based OP-47
standard is in use.
The DVB standard (EN 300-743) defines a bitmap subtitling format, which
allows for greater flexibility (e.g. enabling non-alphabetic languages such as
Arabic or Japanese to be carried), but obviously at greater bandwidth cost.
25 Region 2 DVDs carry image based subtitles in the VOB file of the DVD.
Region 1 DVDs commonly contain EIA-608 data decidable from the picture
edge. Blu-ray m2ts format supports an image-based type called the PGS
(Presentation Graphic Stream) which are bitmaps. For internet delivery,
again, multiple standards exist, and include both text and/or image-based
30 types. Synchronized Multimedia Integration Language (SMIL) supports both
images and text (although it supports far more than just captions), and Timed
5
Text Markup Language (TTML) is an authoring and transcoding standard
used to repurpose television content for the internet.
It would be advantageous to be able automatically to measure and report the
integrity of, and timing consistency of closed captions with respect to the
video they 5 are associated with, throughout a broadcast system or other
content delivery systems.
In some embodiments, this is done with the assistance of existing audiovisual
fingerprinting and correlation techniques, by introducing an additional
fingerprint component (or separate data channel) which carries captions that
10 have a known temporal linkage to the video fingerprint (i.e. timecode or frame
count). The basic idea is depicted in the simplified diagram shown in Figure 1.
This builds on lip sync technology, but between captions and video rather
than between audio and video. It also differs from lip sync technology in the
fact that audio and video are continuous whereas closed captions are
15 intrinsically irregulary in nature, depending on content.
Figure 1 is a simplistic representation of a broadcast system signal path from
a server 10 which plays out AV content to viewers, one of which is illustrated
schematically at 20. Generalised signal manipulation operations are shown
at 30.
20 At appropriate locations over a test section of the signal path, two or more
fingerprint generators are inserted. These are represented in Figure 1 by
Fingerprinter A and Fingerprinter B, which are disposed on opposite sides of
the signal manipulation operations 30 . Each fingerprint generator produces:
􀁸 Video fingerprint data
25 􀁸 Audio fingerprint data
􀁸 Captions data
The video fingerprint data and audio fingerprint data can be generated in a
wide variety of known ways. Reference is directed, for example to WO
2009/104022 which provides examples of video and audio signature
30 generation techniques.
6
The video fingerprint, audio fingerprint and captions data is provided by any
convenient means (for example an IP network) to a fingerprint comparison
unit 40. This may conduct:
􀁸 Video comparison
5 􀁸 Audio correlation and relative delay (to video)
􀁸 Captions correlation and relative delay (to video)
In what follows, attention will be focused on the generation of captions data
and the comparison of captions data with (usually) video fingerprint data.
Captions data
10 In the SDI domain, closed captions are commonly carried in ancillary data
that is intrinsically linked to the video. But it can be extracted, manipulated or
modified and reinserted – for example by a television standards converter.
For this reason, timing changes to the video can be introduced as the signal
propagates through the broadcast chain. The point of interest is whether, at
15 given point in the processing cascade (typically near the end), the timing of
the captions relative to the video is the same, similar to or different to the
original temporal alignment of the captions (i.e. what the content creator
intended).
In modern broadcast setups, captions are commonly converted to images
20 during the chain before distribution to the consumer– MPEG transport
streams carry image-based captions (and display times) but they can also
contain text-based ones. Preferred arrangements for measuring and reporting
caption presence, delay and jitter would operate irrespective of caption carrier
format (i.e. agnostic of text or images), so that the fundamental question of
25 caption timing errors is the same, whatever format exists at the reference
points.
Comparison of raw closed caption data bytes (e.g. EIA-608, “line 21”
captions) – which are embedded in the video fields – is troublesome, not least
because caption modifications or timing changes do not generally just shift
30 the raw data stream forwards or backwards – i.e. not a simple n-byte delay.
7
To measure the caption timing (delay and jitter), the raw stream is decode as
far as the text strings involved, and the frames at which they start and stop
being displayed on the screen. Figure 2 illustrates the basic approach. (This
is a simplistic representation – in reality the raw caption streams need to be
buffered in order to interpret the raw data and decode 5 lines of text. In other
words, there is obviously a causal relationship here – but Figure 2 illustrates
the essential point.)
The blue (top) and orange (bottom) rectangles here represent lines of text
(not words). The comparison involves taking each line from each caption
10 event in one stream (in the diagram, stream 1) within a given temporal
measurement window, and comparing it with each line in each caption event
in the other stream, within a range that included the same measurement
window, plus and minus a detection region.
A given caption might be a start and stop timecode and multiple lines of text,
15 and in this analysis, each line constitutes a separate event. The idea is, to
decode the caption data to an appropriate representation (for example similar
to Subrip (.srt) ) and treat text lines as events. For example, the srt
specification;
1
20 00:00:44,581 --> 00:00:46,556
(SIMPLE TUNE)
2
00:00:53,381 --> 00:00:55,356
25 Ha-ha-ha.
3
00:00:55,381 --> 00:00:57,636
Very good. So? Are you confident?
30
4
00:00:57,661 --> 00:01:00,276
8
She's gonna smash it, aren't you,
babes? Bring it on!
contains three captions that carry single lines of text – three events. The
fourth caption involves two lines – so 5 two events (but with the same start/stop
times).
For each event line matched, a delay value can be deduced from by
comparing the difference in display times. These might be field numbers or
time codes. These need not necessarily be absolute time references – just an
10 indexing which defines in stream 1 on which field in stream 1 the caption
events of stream 1 start and stop. Similarly with stream 2. Correlation of the
video fingerprints between streams 1 and 2 determines which fields are
corresponding (irrespective of timecodes). This match then constitutes a
reference that allows the relative timing of caption event start/stop times to be
15 compared.
The caption event line matching described above provides a timing
measurement for every line of every caption event in the measurement
window. The number of lines and events is both content dependent and
dependent on the choice of window size, but a window size of say 4 seconds,
20 with a detection range of +/-10 seconds is commensurate with audio and
video media matching and lip sync measurement and typically such a choice
would lead to 0-6 caption lines being matched.
With audio and video, delay errors are generally quasi-static (eyes and ears
are very sensitive to jitter, so systems are generally strict about delay
25 consistency). This is not necessarily true of captions, which may exhibit jitter.
The measurement described for individual caption lines provides a means of
measuring (and hence reporting) caption timings both in terms of fixed (quasistatic)
delays and any jitter that is present.
For text captions (e.g. EIA-608, EIA-708 or OP-47), the matching of lines can
30 be done using known text-text correlation technology.
9
The comparison of image and text based captions is less straightforward. In
fact, even the of image captions in one source against image captions in
another source is not straightforward, as the example in Figure 3 shows. In
Figure 3 there is shown actual captions from broadcast output, being
5 essentially the same AV content (the same episode of an episodic
programme) but broadcast at different times on different channels. In this
case, one can see that simply comparing bitmap headers for example, or
image checksums, would not be helpful, with two caption images having been
converted during the delivery chain to a single caption image.
10 Extraction of the text from the images by OCR, followed by text-text
correlation as described in the previous section is an obvious way forward, as
discussed above.
Preferred embodiments of this invention provide an alternative strategy, which
avoids the processing complexity of full OCR by deriving a caption fingerprint
15 from a length of each word in the caption event, without regard to the identity
of the character or characters forming the word. In the example described
here, the correlation is based on word lengths as a proportion of the text line
they appear in, using these as the fundamental “atomic” units for matching.
In this described example, “word length” means not just the number of
20 characters in each word (although that is a significant factor), but the length a
word takes up as displayed. E.g. “WARMER” is 6 characters long, but as
displayed, is longer than “LIVING” which is also 6 characters long. Although
display lengths do vary by font, the relative word lengths are typically similar.
Display lengths for text-based captions can easily be determined, for example
25 by the use of a pre-calculated look-up table based on font averages. Display
lengths for image-based captions can be determined by a series of steps
which are similar to commonly used initial steps of OCR techniques, but are
far simpler, and involve far less processing effort. It is not necessary to
perform full OCR to obtain the word lengths from a captions image.
30 In one example, the lengths of words in image based captions are recovered
with the following steps:
10
1. Convert all input images to white-on-black (binarise the input
greyscale, count black/white ratio and invert the image if necessary)
2. Sum rows of binarised image, and seek each text line start/stop image
rows based on occupied rows. 5 This determines the number of text
lines in the image, and the pixel start/stop row of each text line
3. Low pass filter the input image
10 4. Estimate italic slant by gradient estimation (for example: 1D Lucas-
Kanade type approach) between adjacent lines – using the pre-filtered
image.
5. Interpolate the input (unfiltered) image to correct for italic slant, and
15 binarise the result
6. Sum columns of each text line in slant-corrected image between
detected start/stop image rows to generate a horizontal histogram.
20 7. Histogram lengths of consecutively un-occupied bins in horizontal
histogram to generate a “gaps histogram”
8. Detect the maxima in the gaps histogram (this is the initial intercharacter
spacing estimate)
25
9. From the maxima, increment bins to detect an unoccupied bin. From
this bin upwards, detect the first occupied bin. This is the lower-end
estimate of the inter-word spacing
30 10. Detect the maximum occupied bin in the gaps histogram – this is the
upper-end estimate of the inter-word spacing
11
11. From the initial inter-character spacing estimate, and the lower and
upper end inter-word spacing estimates, form a threshold to
discriminate between inter-character and inter-word spacings
12. Then re-test the horiz 5 ontal histogram of each line to determine which
gaps (consecutive unoccupied bins) correspond to letter gaps and
which correspond to word gaps. This test also gives the pixel start/stop
positions of letters and words in the detected lines
Figure 4 illustrates these steps as applied to a caption image containing two
10 lines. The input image is shown at 4(A). Also shown at the right of the image
are row histograms of the binarised input image, used to detect line start/stop
rows. Figure 4 shows the identified word rectangles at 4(B). The detected
words are shown overlaid on slant-corrected version of the input image.
Figure 4 also shows at 4(C) the character rectangles identified. Experiments
15 reveal that the words are easier to determine reliably than the letters, and the
primary basis of the matching proposed in this disclosure is the words. The
letters are simply a point of interest here. (Letters can merge and words can
be split (or merged). These are well known hurdles in fully-fledged OCR.)
Shown at 4(D) is a filtered input image (luma (inverted ) 3x3 box filter), which
20 can be used for slope estimation. Shown at 4E are line histograms (column
sums [of (unfiltered, inverted ) luma of input image after italics correction]
between detected line stop/start rows)
Having determined the word rectangles from each line, the line is then
characterised by a set of percentages which represent the percentage that
25 each word is relative to the sum of word rectangle lengths.
At the matching stage, for every video field processed, each caption event
(line) in the measurement window of one stream is tested as a match against
each caption event (line) in the other stream over a range which included the
same measurement window, plus and minus a specified detection range, as
30 illustrated in Figure 2. For two lines, A and B, the match between them is
determined by;
12
􀂦
􀂦 􀂦
􀀐
􀀠
􀀐
􀀠
􀀐
􀀠
􀀐
􀀠 􀀐
􀁵
􀀠
􀁵
􀀠
1
0
,
1
0
1
0
1
100 100
Words
k
B
k
Line B
k
Line A
k
A B
Words
i
Line B
i
Line B
Line B k
Words k
i
Line A
i
Line A
Line A k
k
w
M w w
Width
w Width
Width
w Width
Where Line A
k Width is the length of the kth word in line A in pixels, and
Line B
k 5 Width is the kth word in line B in pixels. (The units are pixels, irrespective
of whether the captions originate from images or from text: in the image case,
the word widths are determined as described above by the simplified image
analysis. If the captions originate from text, the word widths are determined
from a look table of average font display widths)
10 Matches are only sought for lines which have greater than one word, and
comparisons are only made between lines with the same number of words in
them. For each event in the measurement window, the best A B M , match
value is selected – this is effectively a match confidence – and accepted if it is
greater than a specified acceptance threshold, 􀁗 . For each match, a
15 corresponding delay value is calculated. The collection of matched events
within the detection window then allows an average confidence and average
delay to be calculated (and of course other measurements such as min and
max values). A record of the number of captions in the measurement window,
the number of matched events and the number of unmatched events are also
20 made.
Matching is conducted A to B as well as B to A, because these are not
necessarily symmetric. For example, if some events were lost from one
channel, this might still indicate a good match to the other, but not vice versa.
13
Figure 5 below shows an example of caption matching between two textbased
caption fingerprints. These correspond to two different Freeview
broadcasts of the same episode of the BBC’s Eastenders, from 1st Dec 2015.
One is from BBC1 broadcast at 7.30pm, the other from the re-broadcast of
5 the same episode on BBC3 at 10pm. Of course, the original broadcasts
involve DVB bitmap captions, not text, and here a statement of matching textbased
captions relates to the transport streams having been decoded and
non-realtime OCR tools being used to recover the text in .srt format, then
fingerprinting using the .srt as input.
10 Figure 5 shows matching of stream 1-2 and vice versa, with match confidence
at the top, delay value in the centre plot, and the number of caption events in
the measurement window at the bottom. In this example, the measurement
window is 8 seconds wide, the detection range is +/-10 seconds, and the
peak acceptance threshold 􀁗 􀀠 0.95 . The slight x-axis displacement between
15 them is a facet of there being a small delay between the fingerprints, which is
also reflected in the fact the 2-1 match reports approximately -60ms delay,
whereas the 2-1 match reports approximately +60ms delay. (N.B. The video
field rate is 50Hz, thus the field period is 20ms. The smaller fluctuations in
delay shown in Figure 5 appear to arise because the display times decoded
20 from the transport stream in for caption images, e.g. 00:00:53,381 (hours,
mins, sec, ms) either contain minor errors intrinsically (and just get rendered
to the nearest field) or are not decoded with strictly accurate timestamps by
the analysis tool used to extract them.
Note that, some gaps do occur, where there are periods over which no
25 captions are specified.
Figure 6 shows a corresponding set of graphs – for the same two broadcasts
– but where the matching is between text-based captions on the BBC1
broadcast, and DVB image captions on the BBC3 broadcast.
The match confidence here is the average value of A B M , defined above, and
30 although this is slightly less than 1.0 (i.e. slightly less than the corresponding
plot in Figure 5 or text-text matching), it is very close. This is understandable
14
– the text-image matching is a harder problem – and still a viable comparison.
Even if the image analysis step correctly identifies the right word rectangles,
the percentage word widths predicted by the text character look-up may
sometimes be slightly mismatched, in some circumstances because the font
of the image is not 5 known. Furthermore, there are more gaps in this case,
which arise when caption lines involve just one word.
There are numerous simple techniques by which this might be improved by
doing a second pass of the measurement window events – once the match of
multi-word lines has been done – whereby the width of single word events is
10 measured (and matched) according to a percentage of the average line
length in the measurement window, or the longest, or the one with the
greatest number of words, or the temporally closest etc.
The gaps which occur with text-text matching and text-image matching, may
be relevant if it is desired to monitor the ‘user experience’ by qualitative
15 differences, and report whether a caption channel is being delivered with
acceptable quality. When no match occurs – because there are no caption
events (which happens several times in Figure 5 and Figure 6) – this is not a
systemic failure. Similarly, a delay of a field or two – in caption appearance
terms to humans – is probably not a big concern.
20 Having performed an “agnostic” matching of captions, the next task is to map
the results into an acceptable (useful) reporting metric. One way to filter the
matching results is by incorporating a state machine. Putting aside potential
questions about the start-up conditions, a basic strategy is to report ‘good’
when captions are matched, providing the delay is within some user specified
25 threshold. When caption matching returns no matches, a temporal counter is
started, which is incremented every subsequent field for which no matches
are found. If a match is encountered, the counter is set to zero, otherwise, if
the counter reaches a specified limit (say 30 seconds) (and the input
fingerprints do contain caption events), then a ‘fail’ is flagged.
30 Additional state machine inputs may include; the number of unmatched
events, the text colour, text screen position, and in the case of text-based
15
matching; character errors and white space differences. Each of these pieces
of data provides further information about the goodness of the caption match..
There has been disclosed an automatic, format agnostic method for matching
caption streams to determine equivalence and delay relative to the video with
which they are associated. By their nature, the event 5 matching is sporadic, so
a state machine may be used to filter the results and generate a simple but
meaningful output flag.
It should be understood that this invention has been described by way of
example only. Thus, there will be other ways – beyond those described above
10 - of processing a caption event to derive a caption event fingerprint.
Preferably the processing is such that the text of a caption event cannot be
derived from the caption event fingerprint. In preferred arrangements, the
caption event fingerprint is derived from a length of each word in the caption
event, without regard to the identity of the character or characters forming the
15 word. Where the caption event comprises a caption image, the length of
each word in the caption event can be determined in a variety of ways by
analysing the caption image to identify image regions corresponding
respectively with words in the caption; and measuring a horizontal dimension
of each image region. The image regions usually correspond respectively with
20 lines of words in the caption and the length of a word is represented as a
proportion of the length of a line.

CLAIMS
1. A method of monitoring audio-visual content which includes a
succession of video images and a plurality of caption events, each
caption event being associated with and intended to be co-timed with a
respective string of successive images, the method comprising the
steps of:
processing a caption event to derive a caption event fingerprint;
searching audio-visual content to identify a caption event
matching a defined caption event fingerprint;
analysing any matching caption event; and
measuring any caption event error.
2. The method of claim 1 where the measured caption event error is
selected from the group consisting of a missing caption event; a
caption event timing error and a caption discrepancy.
3. The method of any one of the preceding claims, in which the caption
event comprises a plurality of words, each formed from one or more
characters, wherein the caption event fingerprint is derived from a
length of each word in the caption event, without regard to the identity
of the character or characters forming the word.
4. The method of claim 3 in which the caption event comprises a caption
image, wherein the length of each word in the caption event is
determined by analysing the caption image to identify caption image
regions corresponding respectively with words in the caption; and
determining a horizontal dimension of each such caption image region.
5. The method of claim 3 or claim 4, where the caption image is analysed
to identify caption image regions corresponding respectively with lines
of words in the caption and the length of a word is represented as a
proportion of the length of a line, preferably where the length of a word
17
I/We Claim:
is represented as a proportion of the length of a line containing the
word.
6. The method of claim 5, where a measurement window of audio-visual
content is defined containing a plurality of caption events and the
length of a word is represented as a proportion of the representative
line length derived from the measurement window, preferably where
the representative line length is the average line length in the
measurement window or the length of a representative line in the
measurement window, preferably selected as the longest line, the line
with the greatest number of words, or the temporally closest line.
7. A system for monitoring audio-visual content which includes a
succession of video images and a plurality of caption events, each
caption event being associated with and intended to be co-timed with a
respective string of successive images, the system comprising:
at least first and second fingerprint generators operating in a
content delivery chain at respective locations upstream and
downstream of defined content manipulation process or processes,
each fingerprint generator serving to process a caption event to derive
a caption event fingerprint; and
a fingerprint processor serving to compare caption event
fingerprints from the respective fingerprint generators to identify
matching caption events; and to measure any caption event error.
8. The system of claim 7 where the measured caption event error is
selected from the group consisting of a missing caption event; a
caption event timing error and a caption discrepancy.
9. The system of claim 7 or claim 8, each fingerprint generator serving to
record the timing of that caption event and the fingerprint processor
serving to determine the timing relative to the succession of video
images of each matched caption event.
18
10. The system of any one of claim 7 to claim 9, in which the caption event
comprises a plurality of words, each formed from one or more
characters, wherein the caption event fingerprint is derived from a
length of each word in the caption event, without regard to the identity
of the character or characters forming the word.
11. The system of claim 10 in which the caption event comprises a caption
image, wherein the length of each word in the caption event is
determined by:
analysing the caption image to identify caption image regions
corresponding respectively with words in the caption; and
determining a horizontal dimension of each such caption image
region.
12. The system of claim 10 or claim 11, where the caption image is
analysed to identify caption image regions corresponding respectively
with lines of words in the caption and the length of a word is
represented as a proportion of the length of a line, preferably where
the length of a word is represented as a proportion of the length of a
line containing the word.
13. The system of claim 12, where a measurement window of audio-visual
content is defined containing a plurality of caption events and the
length of a word is represented as a proportion of the representative
line length derived from the measurement window.
14. The system of claim 13, where the representative line length is the
average line length in the measurement window or the length of a
representative line in the measurement window, for example the
longest line, the line with the greatest number of words, or the
temporally closest line.
19
15. The system of any one of claim 7 to claim 14 where a plurality of
measured caption event errors are combined to generate a flag
indicating whether or not captions are acceptable.