A Convolutional Neural Network for Image Super-Resolution Using Internal Dataset

Shanshan Zhao,Jing Liu,Shancang Li,Xiaoyan Zhang,Yuxin Xue

doi:10.1109/access.2020.3036155

Abstract

Deep convolutional neural networks have recently achieved dramatic success in super-resolution (SR) performance in the past few years. However, the parameters of the mapping functions of these networks require an external dataset for training. In this paper, we propose a convolutional network for image super-resolution reconstruction that can be trained using an internal dataset constructed using a single image. The proposed single image convolutional neural network (SICNN) is designed with two branches. First, a large scale-feature branch trains the feature mappings that are from the low resolution (LR) image patches to the high- resolution image (HR) patches. The LR image patches are the enlarged image patches via bicubic interpolation. Second, the small scale-feature branch trains the feature mappings that are from the down-sampling image patches to the enlarged image patches. In contrast to the existing SR networks, the SICNN enjoys two desirable properties: 1) it does not require external datasets to conduct training, and 2) it enlarges an SR image at an arbitrary scale while restoring the clear edges and textures. The results of evaluations on a wide variety of images show that the proposed SICNN achieves advantages over the state-of-the-art methods in terms of both numerical results and visual quality.

Highlights

Super-resolution (SR) reconstruction seeks to restore high resolution (HR) images from one or more low resolution (LR) images
The major contributions are summarized in three folds: 1) We propose a convolutional network (SICNN) to reconstruct HR images with an arbitrary scale factor using an internal dataset rather than an external training dataset; 2) We propose using dual branches to extract the image features from two scale image patches to improve the performance of the SR image reconstruction (SICNN); and 3) We propose first enlarging the LR image to the proper size, capturing the features and reconstructing its HR image
In this paper, we proposed a novel multiscale network, which creates an internal dataset using the test image itself and exploits a double-branch structure to capture and train the image features at different scales, for SR image reconstruction (SICNN)

Summary

Introduction

Super-resolution (SR) reconstruction seeks to restore high resolution (HR) images from one or more low resolution (LR) images. Super-resolution (SR) reconstruction demands that the image details can be still retained clearly while the reconstructed images are being zoomed in or out on. Superresolution (SR) reconstruction is widely applied to video supervision, medical imaging, military reconnaissance, and remote surveillance [1]–[3]. This technology has attracted great attention over the last three decades and many methods have been proposed. These methods can be divided into the following three categories: interpolation-based methods, reconstruction-based methods, and learning-based methods. Interpolation-based SR methods [4], [5] estimate the missing pixel in an HR image using the neighborhood pixels in the

Objectives

Methods

Conclusion