Caffe_VDSR

This is an implementation of "Accurate Image Super-Resolution Using Very Deep Convolutional Networks" (CVPR 2016 Oral Paper) in caffe.

Instruction

VDSR (Very Deep network for Super-Resolution) is an end-to-end network with 20 convolutional layers for single image super-resolution. The performance of VDSR is better than other state-of-the-art SISR methods, such as SRCNN, A+ and CSCN (My implementation of CSCN).

Update

Multi-Scale Training

Add multi-scale implementation (you can test different scales super-resolution in only one model). The model trained by myself yields similar performances with original paper!

Data Augumentation

Add codes of data augumentation.

Solver Setting

Use Adam instead of SGD. 80 epochs is enough.

VDSR Official Model

Recommend to use "VDSR_Official.mat" if you just want to do some test.

Dependencies

Train

• Caffe

Test

• MatConvNet

Usage

Train

1. Place the "Train" folder into "($Caffe_Dir)/examples/", and rename "Train" to "VDSR"

2. Download training data from 291

3. Open MATLAB and direct to ($Caffe_Dir)/example/VDSR, run "data_aug.m" to do data augumentation, and then run "generate_train.m" and "generate_test.m" to generate training and test data (Code from SRCNN)

4. To train VDSR, run ./build/tools/caffe train --solver examples/VDSR/VDSR_solver.prototxt

5. After training, run "caffemodel2mat.m" to convert caffemodel to mat for testing (matcaffe is required)

Test

1. "Demo_SR_Conv.m" is a simple test code. Just run it and you will get the result

2. "VDSR_Adam.mat" is a model trained by myself

3. "VDSR_Official.mat" is an official model converted from Official Test Code

Performance in PSNR

Factor 2

Factor 3

Factor 4

References

Please cite [1] and this repository if you use this code in your work, thank you!

[1] Jiwon Kim, Jung Kwon Lee and Kyoung Mu Lee, "Accurate Image Super-Resolution Using Very Deep Convolutional Networks", Proc. of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016