ComCNN-RecCNN

The code of paper "An End-to-End Compression Framework Based on Convolutional Neural Networks". TCSVT

Framework

Requirements

• Windows10
• Matlab R2015b
• MatconvNet 1.0-beta23 (https://www.vlfeat.org/matconvnet/)
• CUDA 8.0 (CPU is OK)

How to Run

Training

We have provided the pretrained model of RecCNN and ComCNN in the folder ComCNN/RecCNN_model and RecCNN\data\model.

• Generating the training data of RecCNN model (The pretrained ComCNN or the newly produced ComCNN model is used).
• Training the RecCNN model, and copying the produced RecCNN model into ComCNN/RecCNN_model
• Generating the training data of ComCNN model.
• Training the ComCNN model, (the newly produced RecCNN model is used during training process) and copying the produced ComCNN model into RecCNN\data\model.
• Repeat the above four steps for several times until stabilization of the model.

Testing

Through the training stage, the newly produced model of ComCNN and RecCNN are obtained.

• Executing the test code of ComCNN ComCNN/Demo_Test_Qp_30.m to produce the compact representation of input image.
• Executing the test code of RecCNN RecCNN/Demo_Test_QP_30.m to output the final result of reconstructed image.

Experimental Results

• Subjective results

• Objective results

Additional instructions

• For training data, you can choose any dataset by yourself.
• The code is implemented in terms of JPEG encoder with Qp=30, you can set the configration as you wish.
• The image compresser BPG and JPEG2000 are provided, you can modify the image compresser by yourself (JPEG is used in my repo).
• If you like this repo, Star or Fork to support my work. Thank you.
• If you have any problem about this repo, please email wxcui@hit.edu.cn

Acknowledgments

This code is built based on the repo https://github.com/cszn/DnCNN