Deep Learning Super-Resolution

The goal of this project is to create a deep neural network for single image super-resolution. To solve this problem, we implement a state-of-the-art network (efficient sub-pixel convolutional neural networkESPCN) and build our own from that one (efficient deep sub-pixel convolutional neural network EDSPCN).

Results

Input image :

ESPCN output image :

EDSPCN output image :

Requirements

• Python 3
• TensorFlow
• NumPy
• SciPy

Usage

• Training : python3 train.py

  • --model model : model to train, possible values : espcn, edspcn (edspcn by default)
  • --dataset dataset : training dataset (data/General-100 by default)
  • --batchsize batchsize : size of batch (20 by default)
  • --epochs epochs : number of epochs (1000 by default)

• Super-Resolution : upscale.py

  • --model model : model to use, possible values : espcn, edspcn (edspcn by default)
  • --image image : input image to upscale

• Summary of the training stage : tensorboard --logdir models/save/*MODEL*/train