This is a implementation of "Accurate Image Super-Resolution Using Very Deep Convolutional Networks"[http://arxiv.org/abs/1511.04587] in caffe.

VDSR (Very Deep network for Super-Resolution) is an end-to-end network with 20 convolution 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.

Train: Caffe

Test: Matconvnet

Train:

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

2. Open MATLAB and direct to ($Caffe_Dir)/example/VDSR, run "generate_train.m" and "generate_test.m" to generate training and test data.

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

4. Set clip_gradients in VDSR_solver.prototxt to solve gradient explosion problem

5. Change the learning rate when the error plateaus

Test: