NPEG - Neural Network Image Compression

Encode/decode natural images using Neural Network.

Status

Work in progress

Try it out

$ cd npeg
$ ipython -i train.py

then type in one of the following commands:

# strong noise training (should converge to 0.0008 loss)
for i in range(5):
    r(cnoise=15.0)
    
# weak noise pretraining (no difference in final loss)
for i in [0.1, 0.3, 1.0, 5.0, 15.0]:
    r(cnoise=i)
    
# save weights to file
save()

# load weights from file
load()

# test model on randomly sampled CIFAR
show()

Dependencies:

• TensorFlow r1.0
• Canton library: pip install canton
• Keras (mainly for downloading CIFAR)

If you need visualization functionality:

• cv2 (install via conda is recommended)
• cv2tools (pull from https://github.com/ctmakro/cv2tools then pip install -e <dirname>)
• or modify the code to use other visualization libraries you prefer

How it works

中文读者可以直接看我的博客，有配图：https://ctmakro.github.io/site/on_learning/npeg/npeg.html

For English readers:

Training:

• image -> Encoder CNN -> features
• features += gaussian noise
• features -> sigmoid -> code
• code -> Decoder CNN -> reconstruction
• loss = mean((image-reconstruction) ** 2) + mean(code**2) * 0.01

To reduce reconstruction loss, the best encoding strategy for the encoder is to drive its output ("features") large, to reduce artifacts caused by the gaussian noise.

Therefore by increasing the magnitude of the gaussian noise, the code will eventually saturate to 0 or 1.

We encourage sparsity of the code (to allow for further compression) by adding a penalty term (mean(code**2) * 0.01), after which the code will tend to include more zeros and less ones.

Testing:

• image -> Encoder CNN -> features
• features -> sigmoid -> binary quantization -> code
• code -> Decoder CNN -> reconstruction

About

Author: Qin Yongliang

License: MIT