Residual RNN

This repo is a Tensorflow implementation of paper Full Resolution Image Compression with Recurrent Neural Networks. This repo also contains a trained Residual LSTM model (saved in save/) on a small dataset (doesn't include dataset). Here is one example shown in the paper

Original image:

Reconstruction:

Requirements

• SciPy
• Tensorflow 1.4+

Usage

Training, put training data in imgs/ folder and run following code for default setting. The model parameter can be modified in model.py. Running time comparison of original and reconstructed images can be seen in eval/. Model file is saved in save/model.

python train.py

Encoding

python encode.py --model save/model --input kodim05.png --iters 10 --output compressed.npz

Decoding

python decode.py --model save/model --input compressed.npz --output compressed.png

Evaluation, code from Tensorflow's official repo

python msssim.py -o kodim05.png -c compressed.png