tyshiwo / MemNet

Code for our ICCV'17 paper "MemNet: A Persistent Memory Network for Image Restoration" (SPOTLIGHT Presentation)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

MemNet

Citation

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

@inproceedings{Tai-MemNet-2017,
  title={MemNet: A Persistent Memory Network for Image Restoration},
  author={Tai, Ying and Yang, Jian and Liu, Xiaoming and Xu, Chunyan},
  booktitle={Proceedings of International Conference on Computer Vision},
  year={2017}
}

Other implementation

[MemNet-tensorflow] by ly-atdawn

[MemNet-pytorch] by Vandermode

Implement adjustable gradient clipping

modify sgd_solver.cpp in your_caffe_root/src/caffe/solvers/, where we add the following codes in funciton ClipGradients():

Dtype rate = GetLearningRate();

const Dtype clip_gradients = this->param_.clip_gradients()/rate;

Training (Taking Super-resolution task as the example)

  1. Preparing training/validation data using the files: generate_trainingset_x234/generate_testingset_x234 in "data/SuperResolution" folder. "Train_291" folder contains 291 training images and "Set5" folder is a popular benchmark dataset.

  2. We release two MemNet architectures: MemNet_M6R6_80C64 and MemNet_M10R10_212C64 in "caffe_files" folder. Choose either one to do training.

     $ cd ./caffe_files/MemNet_M6R6_80C64
     $ ./train_MemNet_M6R6_80C64.sh
    

Test (Taking Super-resolution task as the example)

  1. Remember to compile the matlab wrapper: make matcaffe, since we use matlab to do testing.

  2. We release two pretrained models: MemNet_M6R6_80C64 and MemNet_M10R10_212C64 in "model" folder. Choose either one to do testing on benchmark Set5.

     $ cd ./results/MemNet_M6R6_80C64
     $ matlab
     >> test_MemNet_M6R6_SR
    

    The results are stored in "results" folder, with both reconstructed images and PSNR/SSIMs.

More Qualitative results

Image denoising

Super-resolution

JPEG deblocking

About

Code for our ICCV'17 paper "MemNet: A Persistent Memory Network for Image Restoration" (SPOTLIGHT Presentation)


Languages

Language:MATLAB 56.0%Language:HTML 24.3%Language:C 16.7%Language:C++ 2.7%Language:CSS 0.2%Language:Shell 0.1%Language:M 0.0%