SRCNN-Pytorch

Implementation of SRCNN referred in Image Super-Resolution Using Deep Convolutional Networks.

Original implementation in Caffe can be found here.

Requirements

pytorch>=1.3.0

numpy>=1.17.4

tqdm

tensorboardX

Pillow==6.2.1

skimage==0.16.2

winlaic (which can be found in my repositories, use version in https://github.com/winlaic/winlaic/commit/2facccc5bc51894e40589f53a40e93fbbb4be502)

Useage

Train

Please follow these steps to train the model.

1. Generate the data package.

   As instructed by the authors, images are clipped into 33*33 patches with stride of 14.

   We followed this and save the patches in a npz file.

   You can accumulate images in a directory and use

   src/generate_datapack_PIL.py [/path/to/images] [datapack_name] to generate the data package.

   By default, the scale factor is 3. You can specify scale factor by appending --scale-factor [factor].

   Only the gray scale image is saved by default. If you would like to save in RGB format, add --colored argument at the end of command.

2. Train the model.

   Use src/main.py train --data-pack [/path/to/data/package] --validate-image-path [/path/to/validate/images] to start training.

   The program will test the model on validate images and calculate PSNR. The best model will be saved.

   Remember to specify --scale-factor if you changed it in generating data package.

   AdamW optimizing method and Kaiming Initialization is exploited to accelerate training.

Test

Use src/main.py test --test-model [/path/to/trained/model] --test-image-path [/path/to/images/tobe/tested] --output-dir [/path/to/output/dir] to generate super-resolution images.

--colored can be specified to generate colored image. Cb, Cr channel are up scaled by bi-cubic interpolation.

Warning

It seems that opencv-python utilize a different bi-cubic interpolation method.

Low-resolution images generated by cv.resize(..., interpolation=cv.INTER_CUBIC) cause model not working.