1. resize functions with 'bicubic' as option in python and matlab are different and published papers recently generally use matlab to produce low-resolution image
2. Implement SRCNN via Keras with Theano as backend. For a fair comparison with published works, low-resolution images are produced by Matlab imresize function.

Use pre-trained model

run SRCNN_test.m in “test” folder (training set is Yang91) upscaling factor = 3

Training

Note: more data and better result

1. generate training patches using matlab
2. use Keras with Theano as backend to train SRCNN model
3. convert Keras model to .Mat for testing using Matconvnet

How to train your model?

1. generate training patches
2. run SRCNN.py to produce SRCNN model
3. run load_save.py first, then run save_model.m to produce Matconvnet model

Differences with the original implementation

1. use Adam to optimize the network for fast convergence

Dependencies

Theano, Keras, Matconvnet.

If this code is helpful for you, please cite this paper: "Image Super-Resolution Using Deep Convolutional Networks".

Note

this code is based on Keras-1.