Att-BLSTM_codec

This is a PyTorch implementation of the JOURNAL OF BROADCAST ENGINEERING paper, Video Compression Standard Prediction using Attention-based Bidirectional LSTM.

If you find our project useful in your research, please consider citing:

@article{kim2019codec,
    author =       {Kim, Sangmin and Park, Bumjun and Jeong, Jechang},
    title =        {Video Compression Standard Prediction using Attention-based Bidirectional LSTM},
    journal =      {JOURNAL OF BROADCAST ENGINEERING},
    volume =       "24",
    number =       "5",
    pages =        "870--878",
    year =         "2019",
    DOI =          "https://doi.org/10.5909/JBE.2019.24.5.870"
}

The patent about this has been registered. pdf

Dependencies

Python 3.6
PyTorch 1.4.0

Data

We used 18 YUV files and converted them to
M2V (MPEG-2), H263, 264, MP4 (HEVC), BIT (IVC), WEBM (VP8), JPG, J2K (JPEG2000), BMP, PNG and TIFF.
The list of video and image files are on ADD_video_set.xlsx in Korean.

Training

Use the following command to use our training codes

python main.py

There are other options you can choose. Please refer to utils.py.

Test

Use the following command to use our test codes

python test.py

Options: scenario

We use the proposed algorithm for classifying the codec of bitstreams encoded with particular scenarios.
When the value, training_scenario is,

0 : training samples not encoded
1 : training samples not encoded and bitwise inversed
2 : training samples not encoded and bitwise xor-d
3 : training samples not encoded, bitwise inversed, bitwise xor-d and endian-swaped every 2 bytes
4 : training samples not encoded and endian-swaped every 2 bytes

When the value, test_scenario is,

0 : test samples not encoded
1 : test samples bitwise inversed
2 : test samples bitwise xor-d
3 : test samples endian-swaped every 2 bytes

Another scenario was implemented by mcthehruj.

Results

When we tested using data with 3 codecs, MPEG-2, H.263, and H.264, we made the network with accuracy 99.39%.
When we tested using data with all 11 codecs, we made the network with accuracy 96.09%.
96.09% is the accuracy of the confusion matrix.

Contact

If you have any question about the code or paper, feel free to ask me to ksmh1652@gmail.com.

Acknowledgement

Thanks for graykode who gave the implementaion of Att-BLSTM.

This work was supported by the research fund of Signal Intelligence Research Center supervised by Defense Acquisition Program Administration and Agency for Defense Development of Korea.