Research On Image Super-resolution Reconstruction Algorithm Based On Convolutional Neural Network

With the rapid development of deep learning technology,traditional machine learning methods in the field of image super-resolution can no longer meet people's increasing demand for accuracy,efficiency,speed,etc..So convolutional neural network that can solve the existing dilemma is presented.The core idea of image super-resolution reconstruction refers to the process of recovering high-resolution images from lowresolution images.It is an important image processing technology in computer vision.Image super-resolution reconstruction is a classic ill-posed inverse problem,because there are multiple solutions for any low-resolution image input.And image superresolution reconstruction is a research difficulty in the field of image processing.Since the introduction of convolutional neural networks,more and more excellent variants of convolutional neural networks have been proposed one after another,and have achieved good results in their different applications.Compared with the traditional machine learning methods,the previous various CNN variants or methods have many advantages.But there is still many improvement on the visual perception and the accuracy of reconstructed images.In view of the above-mentioned problems,the design work of this paper is as follows:(1)This paper introduces the concept of gating mechanism into convolutional neural networks,and proposes a neural network with pre-training strategy-Gated Convolutional Neural Network(PGCNN)for super-resolution of remote sensing images.PGCNN consists of several residual blocks with long jump connections.Each residual block contains a well-designed gated convolution unit,which provides different weights for high-frequency information and low-frequency information to control the transmission of information.The advantage of PGCNN is that the main network can focus on learning more high-frequency satellite image feature information.The experimental results on the four public remote sensing test data sets of SIRI-WHU,NWPU-RESISC45,RSSCN7 and UC-Merced show that compared with several existing classic methods,this method has a comprehensive improvement on accuracy and visual effects.(2)This paper combines the local attention mechanism with the residual structure,and designs a type of residual block-the local attention mechanism residual block(RLA).It can not only effectively use the powerful learning ability of the residual structure,but also use the local attention mechanism to constrain the residual learning process.RLA can learn more directly the local area information of pixels that are more important for image super-resolution tasks.By inserting the RLA model,a single-image superresolution reconstruction network model is proposed,named Local Attention Convolutional Residual Network(RLAN).The experimental results on the five natural image data sets Set5,Set14,BSD100,Urban100 and Manga109 show that compared with several excellent super-resolution methods,the proposed method can recover more detailed information.And the recovered image has more natural texture,subjective visual effects and objective quality.This paper compares the two proposed image super-resolution reconstruction methods with a number of classic image super-resolution methods.Both the subjectively evaluated visual effects and the objectively evaluated reconstruction quality have shown excellent performance,which can be applied to a variety of real scenes such as natural images and satellite remote sensing images.