Matconvnet Training Code for VDSR

A DagNN Matconvnet training implementation of "Accurate Image Super-Resolution Using Very Deep Convolutional Networks," CVPR, 2016.

Coding Environment

• Matconvnet 1.0-beta23 from http://www.vlfeat.org/matconvnet/ installed and compiled.
• Matlab R2016a.
• Titan X Pascal.

Files

• matconvnet-1.0-beta23: Copy this folder to your installed Matconvnet path. Includes a pdist layer function for DagNN.
• cnn_train_dag_hardclip.m: A cnn_train_dag variant specifically for VDSR's gradient clipping.
• create_traindata.m: Creates an imdb file for training dataset.
• exe_test.m: A function for testing test images.
• extract_subim.m: A subfunction for extracting subimages from images. Used in the create_traindata.m file.
• main_train.m: A main code for training VDSR. Calls vdsr_setup.m and exe_test.m.
• vdsr_setup.m: A function for setting network structure for VDSR. Calls cnn_train_dag_hardclip.m