/super-resolution-pytorch

Primary LanguagePython

(WIP) Deep Learning-based Single Image Super Resolution

Implemented with PyTorch

Introduction

import net
srcnn_ex = net.srcnn(f2=5)
print(srcnn_ex)
import utils
print('total parameters: {}'.format(utils.count_parameters(srcnn_ex)))

The output should given:

Sequential (
  (0): Conv2d(1, 64, kernel_size=(9, 9), stride=(1, 1), padding=(4, 4), bias=False)
  (1): ReLU (inplace)
  (2): Conv2d(64, 32, kernel_size=(5, 5), stride=(1, 1), bias=False)
  (3): ReLU (inplace)
  (4): Conv2d(32, 1, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), bias=False)
)
total parameters: 57184

TODO

  • Add training and testing scripts
  • Add more models
  • Add function for visualizing the networks

Models

  • SRCNN
  • FSRCNN
  • DRCN
  • VDSR
  • SubPixel
  • SRGAN (SRResNet)
  • EnhanceNet
  • LapSRN

References

[1] Chao Dong, Chen Change Loy, Kaiming He, Xiaoou Tang, Learning a deep convolutional network for image super-resolution, ECCV, 2014.

[2] Chao Dong, Chen Change Loy, Xiaoou Tang, Accelerating the Super-Resolution Convolutional Neural Network, ECCV, 2016.

[3] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, Wenzhe Shi, Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, arXiv, 2016.

[4] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang, Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution, CVPR, 2017.