Super Resolution Convolutional Neural Network (SRCNN)

SRCNN implementations for Python/Torch, Numpy and Avnet's ZedBoard

The aim of single image super-resolution is to recover a high-resolution image from a single low-resolution image.

This repository includes three implementations of the Super Resolution Convolutional Neural Network (SRCNN) by Dong, Chao, et al. "Image super-resolution using deep convolutional networks." (https://arxiv.org/abs/1501.00092v3)

Summary

• 9-5-5 Convolutional Neural Network
• Three implementations
  • Torch: For training the network, extracting weights and upscaling
  • Numpy: For upscaling
  • Cython: For upscaling (able to be run on Avnet's ZedBoard)
    • Purely CPU-run or
    • Running the 2D convolutions on the ZedBoard's FPGA

Usage

See the README.md files of the respective implementation in their folders for usage instructions and prerequisites:

• Torch
• Numpy
• Cython

Metrics

Average of each metric for the Set5 dataset.

Metric	Bicubic	SCA	NE+LLEA	KKA	ANRA	A+A	SRCNN (Dong)	SRCNN (Ours)
PSNR	29.56	31.42	31.84	32.28	31.92	32.59	32.75	31.92
SSIM	0.871	0.882	0.896	0.903	0.897	0.909	0.909	0.913
IFC	3.49	3.16	4.40	4.14	4.52	4.84	4.58	4.41
NQM	27.93	27.29	32.77	32.10	33.10	34.48	33.21	33.04
MSSSIM	0.975	0.980	0.984	0.985	0.984	0.987	0.987	0.987

Examples

• Ground truth is the original (ideal) image for comparison with the upscaled versions.
• Bicubic is the ground truth downscaled by a factor of 3 and upscaled by a factor of 3 using simple bicubic interpolation.
• SRCNN is the same as Bicubic but using the trained convolutional neural network.

Ground Truth	Bicubic x3	SRCNN x3