Google Summer of Code Update – July 2017

I’ve been working for Red Hen Lab and CCExtractor as part of GSoC 2017. This year, my project was divided into three major components:-

  1. TV Commercial Classification – Details and code here (Done during June)
  2. Improving the visual recognition pipeline
  3. Optimizing the performance of the visual recognition pipeline and deploying it

Apart from this, I have also been working on analyzing and fixing issues with the CCExtractor code in order to better cater to some of Red Hen’s new international use cases.

In this post, I will describe what I have been up to this month, what things I have been spending my time working on, and some interesting decisions that I have had to make based on what I have observed in my experiments.

My Current Setup

Red Hen’s GSOC students have been working extensively on solving machine and deep learning problems which typically require a GPU for computational tractability. The Case HPC is where most of the important number crunching happens. However, being a shared resource, the HPC (and especially it’s GPU nodes) were not always readily available. I recently purchased a new laptop with a 4GB Nvidia GTX 1050 Ti GPU, which I have been using to work locally (as well as on my home institute cluster with more GPUs). Most of the benchmarking work that I have done have been on my computer, however the linearity of the observations and the computational power of other GPUs should hold for our particular use case of image classification (ie a more powerful GPU should take a linearly shorter time and a less powerful GPU a linearly longer time for the same use case).

Benchmarking ResNet for News Shot Classification

AlexNet (CaffeNet in our implementation) is the deep neural network architecture used for the shot characterization into 5 classes in the pipeline. This is a 7 layer deep architecture.

ResNet is a newer, deeper and more accurate model developed by Microsoft Research. It achieves state-of-the-art results on large scale visual recognition and is in essence a modern day upgrade to the 2014 CaffeNet architecture that we use.

Upon training ResNet on the same task with the same training/testing data split to identify news shot categories, the saturation accuracy of ResNet was superior. It achieves a peak accuracy of 93.7% whereas the AlexNet model achieves a peak accuracy of 88.3%. The ‘real’ accuracy (on exhaustive testing) for AlexNet is close to 86.5% and that for ResNet is close to 91%.

The training graph for the two architectures (done on a K40 GPU and a 1050 Ti GPU respectively) looks like this:-


ResNet takes a slightly larger time to converge, likely because of the deeper architecture requiring more time for the gradients to flow across the network. But it achieves an overall accuracy of roughly 5-6% higher than the AlexNet model after training is complete. This is to be expected because the ResNet model is newer and more state-of-the-art. However, it’s extreme depth in terms of layers leads to a much, much higher time taken to train and test (predict new samples using) the model.

While running 100 iterations 10 times and averaging the time values for a batch size of 10 at test time, the benchmarks for my Nvidia GTX 1050 Ti 4GB GPU, and that for an Intel Xeon CPU node look something like this:-

Model Avg GPU Mode time (s)
Avg CPU Mode time (s)
AlexNet 7.34 283.66
ResNet 49.54 1812.31

My GPU seems to be around 3 times slower than a GTX 1080 judging by these benchmarks.

The final decision that I took on the basis of these observations was to stick to the existing AlexNet framework for the sake of speed on CPU nodes. An accuracy increase of 5-6% for the shot type categorization from an existing ~86% was good, but the tradeoff of 6-7 times the runtime was perhaps not ideal for our use case of processing over 300,000 hours worth of video.

Brazilian Timestamps and Deduplication

There was an issue with the Brazilian ISDB subtitle decoder which caused broken timestamps (Timestamps are the primary key for the Red Hen dataset, and without them, the data can’t be part of the searchable archive). There was also an issue with duplication

French OCR – Fixing an Issue with Image Transparency

This one was one of those problems that happened to be a one line fix, but involved a huge amount of analysis, reading and experimentation to get done right. It took me the good part of a week to get this to work as it is now, and hopefully this yields near perfect OCR results for all use cases, including those with transparent DVB subtitles.

Adding GPU specific usage to the Visual Pipeline

Earlier, the pipeline was supposed to be run as one single job on a CPU compute node, and even if the requested node was a GPU node, the capabilities of the GPU would not be used by it. I added the capability to use a GPU if available. If no GPU is available (or detected by the code), we fall back to default CPU execution.

The changes to make this happen can be seen at:-

This is especially useful in speeding up the runtime of the feature extraction and classification steps that involve deep neural networks (namely anything that invovles running a Caffe model). GPU execution of these steps speeds up the runtime by an exponential factor which could be anything between 20 to 200 times depending on the individual computational power of the CPU/GPU in question. On my computer, a nearly 40 times speedup can be observed (benchmarks above).

Upgrading the YOLO person detector in the Visual Pipeline to YOLOv2

The current pipeline has the YOLO object detector specifically for person detection. However, YOLO was upgraded to YOLOv2 this year and has been accompanied by significant accuracy gains. I upgraded the version of YOLO used in the pipeline to the latest one, while retaining the same output format in the SHT and the JSON files.

I integrated the original C code based on Darknet for YOLOv2 to the pipeline, and person detection results are slightly better than before.

Singularity Container for Portable HPC Execution of the Visual Pipeline

Singularity is an HPC friendly alternative to the popular Docker framework. Red Hen has been heavily using Singularity this GSoC. One particularly useful use case is portable usage of the Singularity image on multiple HPC clusters (e.g. Case HPC, one of the many clusters at Erlangen HPC etc)

I have written a basic singularity image for the state of the code at this moment, which creates a container and then downloads all required models and dependencies, and sets up the container for usage of the visual recognition pipeline.

Upcoming Work

In the final third of GSOC, I will work on reducing the overall runtime of the pipeline by as much as possible. I will also work on writing an HPC job manager script which logically segments the different parts of the pipeline and submits and tracks different jobs for each based on available resources (GPU/CPU). Another thing to do is set up the pipeline with CPU optimized Intel Caffe which will allow automatic parallel processing on CPU nodes on HPC. After doing this and testing for sanity, the pipeline should be ready to be put into production on the entire NewsScape dataset, perhaps on multiple HPCs.


Google Summer of Code, Week 12 – GUI Integration and DVB Languages

Google Summer of Code, Week 12 – GUI Integration and DVB Languages

My work for this week focused on integrating the new features that I developed throughout GSOC into the GUI, and also adding support for multiple OCR languages to DVB subtitles.

GUI Integration

The most common users of CCExtractor use the GUI, and may not necessarily be experienced with the command line. Hence, it was important to integrate the new features into the GUI so that they may be available to a much wider audience.

The GUI is essentially a program which provides a graphical interface to the user who can supply the required options with ease, and the program in turn calls the ccextractor executable after parsing the options that the user supplied and converting them into a command string. The GUI interacts with the executable and works just like the normal executable would from the command line, but it gives easy to interpret visual feedback to the user, bundled in an easy to use application.

In addition to the SourceForge Windows GUI, CCExtractor also has a cross-platform Qt GUI which was developed by Oleg Kisselef in GSOC 2015 ( I needed to add parsing support for the options which I had added to the main program and create an interface which would allow those options to be sent to the executable.

It was fairly easy to add the options to the ‘Options’ window. A main checkbox needs to be checked in order to access the other parameters (equivalent to how -hardsubx needs to be specified before any related options on the command line). After I had created the UI, I had to map the elements in the GUI (radio buttons, sliders, checkboxes) to events in the application which would pass on the appropriate commands to the executable. I also verified that the options worked as intended, and constrained them to take only valid values, and also be initialized to the default and recommended values.

The resultant additions to the Qt GUI on Linux look like this:-


DVB Languages

Another thing which I did this week was to add support for potentially 99 different languages using Tesseract’s .traineddata files. Before this point,  only English was supported.

Adding this feature was like solving a partially solved jigsaw puzzle. I just had to complete some of the existing code to search for Tesseract language packs and make sure that it looked for the necessary files in the correct locations.

Initially, I had also added special cases for certain languages like Chinese (simplified) which seemed to come with non standard language codes in the video stream. However, instead of hard-coding a particular case like this, it was deemed better to let the user specify the non standard names, if at all necessary.

I added the -ocrlang and -dvblang parameters. -dvblang allows the user to select which language’s caption stream will be processed. In the event that there were multiple caption streams in the video, only the one specified by the parameter would be processed. -ocrlang allows the user to manually select the name of the Tesseract .traineddata file. This option is helpful if you want to OCR a caption stream of one language with the data of another language. e.g. -dvblang chs -ocrlang chi_tra will decode the Chinese (simplified) caption stream but perform OCR using the Chinese (Traditional) trained data. This option is also useful when the Tesseract .traineddata files don’t come with standard ISO names.

Google Summer of Code, Weeks 10 and 11 – Different Types of Subtitle Detection

Google Summer of Code, Weeks 10 and 11 – Different Types of Subtitle Detection

In some cases, the detected text may be filled with noise and unwanted artifacts. Hence, there was a need to improve the text classifier in order to try and improve the quality of the detected captions. I set up three different levels of Tesseract subtitle line classifiers, and added confidence based thresholds as possible parameters which could potentially improve the quality of the OCR results.

The Three Modes – Frame, Word and Character

There are three different modes at which I used Tesseract to process a particular frame:-

  • Frame: In this mode, the entire frame is processed at once, and the entire UTF8 text detected by Tesseract is written to the caption file.
  • Word: In this mode, every word detected in the frame is individually processed. It can be later thresholded based on confidence or whether it is a dictionary word, or an expletive word etc.
  • Letter: In this mode, every letter detected in the frame is individually processed. It can also be later thresholded. This mode is just present because of the possibility to make it so. For any practical purposes, the first two should serve fine.

I created a parameter called -ocr_mode which allows the user to specify the level at which the OCR will be performed and interpreted. The default is the ‘frame’ mode.

Tesseract Confidence Ratings

The Tesseract Engine supplies confidence ratings along with its OCR predictions. I chose to use these confidence values to improve the quality of text recognition performed by the system. I created an optional parameter called -conf_thresh which allows the user to put a threshold on the confidence rating of the text classification by Tesseract (having a default value of 0, ie all classifications accepted). Only classification results which had a confidence above the threshold were processed and written as captions.

The confidence thresholding works for each of the three OCR modes as described above. For the ‘frame’ mode, the confidence used is the mean text confidence, for the ‘word’ mode, the per-word confidence, and for the ‘letter’ mode, the per character confidence. These results are then thresholded and only the good ones remain.

Italic Detection

Another small part of my proposal was to detect if the formatting of the subtitles was italic. I had originally intended to do this using an orientation estimation using the Fourier transform or instead looking at the average angle of the longest lines in the characters found in their Hough Transform.

However, none of that proved to be necessary since the Tesseract API had a call to detect word font attributes. So, whenever italic detection was to be done, I set the OCR mode to word-wise and called the Tesseract API which would then determine if the word was italic.

An excerpt from the video at (about British warship HMS Bulwark) is:-

00:00:00,000 –> 00:00:04,959
<i>The ship is an enormous machine </i>

00:00:04,961 –> 00:00:07,919
<i>one of the complicated machines that Britains ever built </i>

00:00:07,921 –> 00:00:09,879
<i>we make our own water from sea water </i>

00:00:09,881 –> 00:00:10,959
<i>we deal with our own sewage </i>

00:00:10,961 –> 00:00:12,959
<i>we cook our own food </i>

00:00:12,961 –> 00:00:15,879
<i>we’re a floating city. </i>

Google Summer of Code, Weeks 8 and 9 – Detecting DVB Subtitle Color

Google Summer of Code, Weeks 8 and 9 – Detecting DVB Subtitle Color

DVB Subtitles

DVB (Digital Video Broadcasting) is the standard for TV video in a large number of countries, and is especially prevalent in Europe. In a DVB video stream, subtitles are present as colored bitmap images, which are simply overlaid on the video if subtitles are turned on in the viewing system.

CCExtractor already had excellent support for DVB subtitle text recognition, using Tesseract. It was done by first binarizing the bitmap so that text and the background were separate. This resulted in accurate text recognition by cleaning up the image, but lost color information in cases where multiple colors of text were present in a single bitmap. An additional requirement was to detect the color of each word in the subtitle.

Why Color Is Important

Color changes in DVB subtitles refer to speaker changes in the program. Assigning a different color for a different speaker enriches the assistive capabilities of captions (e.g. for hearing impaired people). Speaker change detection also holds a very large importance for various text processing algorithms for which CCExtractor is a major data source.

Bitmaps and Color Histograms

Bitmaps are just 2-D arrays of numbers, along with an accompanying palette. A palette is like a dictionary which represents a mapping from the pixel value in the bitmap to the actual RGB value. For example, the bitmap may have values ranging from 1 to 8, and 1 may represent Black (0,0,0), 8 may represent White (255,255,255) and so on. DVB subtitles are also bitmaps with their corresponding palettes. Color Histograms are a way to represent the frequency of each color in the image. They are a frequency representation of every single pixel value in the bitmap. The more the amount of a particular pixel value in the image, the higher will its histogram value be.

Word-Wise Color Quantization

The color detection for every word is done by iterating over the bounding boxes of every word obtained in the original DVB OCR results. For every bounding box, a color quantization process is performed. Color quantization essentially means changing the pixel value to a nearby value which has a much higher frequency in the histogram. Using this information from a two bin color quantization, the background and foreground colors are determined, and the foreground (text) color is assigned as the detected color of the text.

Successive words with the same color are grouped together and the points at which the color changed are marked with <font> tags.


A DVB frame with three colours along with the corresponding output is as shown:-


00:01:47,780 –> 00:01:51,339
<font color=”#00ff00″>So he spent last night in a cell?</font>
<font color=”#ececec”>It’s a ROOM. Not a cell. </font><font color=”#ffff00″>Ian!</font>

The color values are exactly what their pixel values are in the bitmap.

DVB Crash Fix

As a result of working on DVB Color Detection, I also noticed and fixed an important bug which was causing a lot of periodic crashes while continuously processing DVB subtitles. The bug was largely due to Tesseract OCR returning multiple newlines at the end of a line. I made a quick fix by increasing the memory allocated to the resulting string variable. It resulted in a large increase in the stability of the DVB processing pipeline.
Although there are still a few issues and bugs in the program, the DVB system is quite stable.

Google Summer of Code, Weeks 6 and 7 – Detecting Colored Subtitles

Google Summer of Code, Weeks 6 and 7 – Detecting Colored Subtitles

Till this point, I have a system which works well for burned-in white subtitles and generates a timed output file. The next step is to add the same support for colored subtitles too.

The HSV Color Space

The HSV color space, and the Hue component (H) in particular, is an excellent representation of the exact color value of a pixel. The normal RGB space requires 3 values to represent the color, whereas the H component takes a value in the range of 0-360 and gives the necessary color information.

You can read more about the color space here.

The chart below shows how the values of H vary for different types of colors.


I exploited a conversion from the RGB to the HSV space in order to detect colored subtitles. Just like there was a luminance threshold in order to detect white subtitles, there is a threshold around the range of the user-specified hue value in order to detect subtitles of a particular color.

This hue based thresholding, along with the existing vertical edge dilation was used to detect subtitles of a particular color.

Color options in the program

The program has 7 predefined color names. The first and most prevalent case is White, the detection of which is luminance based. The other 6 are equally spaced in the hue value range. The colors, along with their hue values, are:-

  1. Yellow – 60
  2. Green – 120
  3. Cyan – 180
  4. Blue – 240
  5. Magenta – 300
  6. Red – 0

Each of these colors can be specified along with the -subcolor option. For example:-

ccextractor video.mp4 -hardsubx -subcolor yellow

In addition to these preset values, there is also the possibility to supply a custom hue value. This value is a custom value between 0 and 360 (not included) which can be supplied to the subcolor option, and could be of help to users who want to extract subtitles of the precise hue value in their stream if it fails to meet one of the presets.

Local Adaptive Thresholding

In addition to detecting colored subtitles, I was also able to improve the detection of white subtitles using local adaptive thresholding algorithms, and Sauvola Binarization in particular. This was an additional step which marginally improved the quality of results for white subtitles (which always have a pixel value greater than their surroundings), however could not be applied to colored subtitles in all cases due to a wide variety of contrasting backgrounds.

Google Summer of Code, Weeks 4 and 5 – Determining Subtitle Appearance Time

Google Summer of Code, Weeks 4 and 5 – Determining Subtitle Appearance Time

So far, I have been able to successfully extract white colored subtitles at an interval of 25 frames, and the output looks decent. However, I need to now actually created a timed transcript (e.g. an SRT file).

Original Plan

I had originally intended on having two strategies to determine subtitle time, which I had described in my proposal as:-

  1. A linear search across the video at a certain interval. Whenever a subtitle gets detected, a binary search will be performed in a window around that frame. Using this, we will detect the exact time of the beginning and the end of the particular subtitle line. This will be of benefit in sequentially processing a file (possible use case of processing a live stream as it is being recorded).
  2. If the entire video is already available to us, instead of doing linear search which will involve a lot of processing overheads for frames in which there are no subtitles, we can directly do a binary search on the entire video to detect subtitle lines. We will get the exact timing of the line as described above, but the overall processing will be faster

However, neither of them were possible, due to constraints which I had not originally anticipated.

FFMpeg Constraints

It turns out that binary search was not a viable option at all, because I could not arbitrarily seek to a timestamp in a video using the FFMpeg library. The closest thing which I could do was seek the file to the nearest I-frame and then iterate through frames to the desired timestamp and then reconstruct the needed frame. However, in a binary search, the whole point of which was to optimize the search, this way would create a massive processing overhead and high redundancy in reconstructing frames during the search. Instead, a linear search with a specified step-size seemed a much better option.

The problem that I described is fairly well documented online:-

New Plan – Efficient Linear Search

I decided to use a linear search across the video with a specified step-size, which was a parameter called the minimum subtitle duration. I set the default value for this as 0.5 seconds, which seems a reasonable assumption for most subtitles.

I also needed to convert times to a single format (milliseconds), from the various different time bases that various different video streams could have. From here, I iterated through the video and sampled frames at regular intervals. The decision that a subtitle line was the same as the last encountered one was when it’s Levenshtein distance was very low. This was necessary in order to combine successive detections which were off by a character or two, which happens quite often due to the natural noise present in the video stream. Whenever the detected subtitle line ended, I would encode it with the seen times.

Integrating with the CCExtractor Encoder

It was really easy to integrate the calculated time with the CCExtractor encoder structure (which with itself brought full output parameter functionality). All I had to do was call two functions at the appropriate times in my code:-

add_cc_sub_text(ctx->dec_sub, subtitle_text, begin_time, end_time, “”, “BURN”, CCX_ENC_UTF_8);
encode_sub(enc_ctx, ctx->dec_sub);

That says a lot about how well written and modularized the existing library is.

And oh, I chose the subtitle mode ‘BURN’ myself. It stands for burned-in subtitles xD.

Example Output

The SRT output for the video at (A Gerard Pique interview), looked as follows:-

00:00:00,000 –> 00:00:06,919
Well, they’re both spectacles,

00:00:06,921 –> 00:00:08,879
NBA basketball as well as football here.

00:00:08,881 –> 00:00:12,879
It’s a spectacle for the fans, they enjoy it and see

00:00:12,881 –> 00:00:13,919
something different.

00:00:13,921 –> 00:00:19,919
And I think a player like Messi could be compared to Curry

00:00:19,921 –> 00:00:21,919
in the USA,

00:00:21,921 –> 00:00:24,879
because they have created something special

00:00:24,881 –> 00:00:26,999
something not seen before,

00:00:27,001 –> 00:00:30,919
and something that makes people excited, ecstatic.

It looks pretty good, and the times are pretty close to perfect, with some variation at the extremes due to those edge frames not being processed. A lower value for the minimum subtitle duration will give even more accurately timed results, but will take a longer processing time.