README

Requirements:

• Python 3+ (download: https://www.python.org/)
• TensorFlow (pip install tensorflow in command prompt)
• Keras (pip install keras)
• OpenCV (pip install opencv-python)
• NumPy (pip install numpy)
• Dahuffman (pip install dahuffman)
• ffmpeg (follow steps: https://www.wikihow.com/Install-FFmpeg-on-Windows , https://extract.me -- to convert 7z file to zip)

Steps: All these files are in the code folder

[NOTE: Running the files might give a warning (deprecation due to compatibility issues of TensorFlow 2.0 library and TensorFlow 1.0 code)]

1. Train the auto-encoder (Optional, since I pretrained the model, takes a lot of time)
   • run train.py
2. Encode your video file [Huffman Encoding]
   • run encoder.py
3. Decode your video file [Decoding and Stream Generation]
   • run decoder.py
4. Control the bit rate to match the compressed file [Generating a better quality lower bandwidth stream than plain H.264 video]
   • ffmpeg -i Football_decoded.mp4 -vcodec libx264 -y -b:v 3800k Football_final.mp4 (run this command in command prompt (set location of prompt to data folder first) )

Output files explained:

• Results Directory:

  • Images from various stages of execution

• Saved_Model directory:

  • decoder_weights: weights generated for decoder after training autoencoder [nupmy array file]
  • encoder_weights: weights generated for encoder after training autoencoder [nupmy array file]
  • huffman_codec: byte stream of Huffman encoding done by encoder.py

• Data directory:

  • Football_raw.mp4: raw mp4(1920*1080,30fps) footage
  • Football_compressed.mp4: Plain H.264 compressed video (1920*1080,30fps)
  • Football_predicted.mp4: Residual predicted by model for the video
  • Football_decoded.mp4: The framewise plain addition video of compressed and residual footage
  • Football_final.mp4: Bit rate adjusted final compressed video generated
  • Football_residual: the residual generated [nupmy array file]