Deep Learning Image Super Resolution

What is the Project About

• This is a deep learning project based on the Image Super-Resolution Using Deep Convolutional Networks - SRCNN paper using the PyTorch deep learning library.

Framework and Dependencies

• The project is built on PyTorch 1.4.
• You will need MATLAB to execute the .m files.

Directory Structure

• The following is the directory structure to arrange everything for the project.

  ├───input
  │   ├───bicubic_2x
  │   ├───bicubic_4x
  │   ├───bicubic_rgb_2x
  │   ├───bicubic_rgb_4x
  │   ├───General100
  │   ├───Set14
  │   ├───Set5
  │   ├───T91
  │   ├───T91_G100
  |    train_mscale.h5
  ├───outputs
  └───src
  

• input: contains the datasets that are used for training and testing. The train_mscale.h5 is the training datasets that gets generated after running the generate_train.m file.

  • Currently the model has been trained on both T91 and General100 image datasets. Both of these datasets are merged into T91_G100 folder. The same corresponds to in the generate_train.m file.
  • The bicubic_x folders contain the blurred images that we use for testing. Generate those images using the bicubic.py file inside the src folder.

• The outputs folder will contain all the output files along with the trained model.

• src contains the python and MATLAB files.

Note: I have take the MATLAB codes from the SRCNN-Keras repository. The original generate_train.m file generate greyscale sub-images. I have formatted the code so as to generate colored (RGB) sub-images. As such, in this project, you will be able to train a neural network model that can carry out super-resolution on RGB images. Please go through the code for more details.

Dataset

• You will find the datasets used in this project and more super-resolution datasets here.

Execution

• generate_train.m: To generate the train_mscale.h5 sub-images.
• Execute the python scripts while being within the src folder in the terminal.
  • python bicubic.py --path ../input/Set14 --scale-factor 2x: To create low-resolution bicubic images for the Set14 data by a scaling factor of 2x.
  • python train.py: For training the SRCNN model.
  • python test.py --input ../input/bicubic_rgb_2x: To generate high resolution images of the 2x scaled low-resolution images.

Some Results

• The following are from testing on the 2x scaled low-resolution images.

References

• Image Super-Resolution Using Deep Convolutional Networks - SRCNN.

• SRCNN-Keras: For the generate_train.m file to create the train_mscale.h5 training data.

• SRCNN-Tensorflow: For the test images.