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

Image super-resolution technology has a very wide range of applications and very high research value in many fields such as medical image processing,remote sensing image processing,and surveillance video processing.Image resolution technology is essentially an ill-posed problem.The mapping relationship between low-resolution images and high-resolution images is not unique.Therefore,the problem of image super-resolution is also a difficult point in the field of image restoration research.In recent years,the super-resolution technology based on convolutional neural networks has shown great advantages and has received extensive attention and research and has gradually become the mainstream image super-resolution method.In order to obtain clearer super-resolution reconstructed images,more and more huge network models are proposed.But the deeper and wider the network is,the more difficult it is to train.At the same time,the size and reasoning speed of the network severely restrict its application in real-world scenarios.In view of the above problems,the main work are as follows.(1)This paper proposes a neural network applied to remote sensing image super-resolution,called Progressive Residual Deep Neural Network(PRDNN).PRDNN uses a new type of progressive residual structure,which can gradually discover satellite image feature information under different levels and different receptive fields,and deeper residual units can provide larger receptive fields,thereby reconstructing super-resolution satellites the image provides more detailed features.(2)This paper proposes an extended residual structure and uses the residual structure to construct a new model,named ERSR.Compared with the traditional residual structure,the extended residual expands the number of output channels before the non-linear activation unit,and a convolutional layer with a convolution kernel size of 1 is used at the end of the residual.This structure can make more effective information can be retained in the residual learning process and improve the efficiency of residual learning.And because it only uses a convolutional layer with a convolution kernel size of 3 and a convolutional layer with a convolution kernel of 1,it will not bring additional parameters,and it can also maintain a very fast inference speed.(3)This paper uses a larger batch to train the network model in the ERSR training process.Experiments show that this strategy can not only greatly improve the convergence speed of the network,but also improve the accuracy of the network after the final convergence to a certain extent and prevent the network from falling into a local optimal value and failing to fully converge.This paper compares PRDNN and ERSR with mainstream image super-resolution methods in recent years,including objective image quality evaluation methods and subjective image quality evaluation methods such as peak signal-to-noise ratio,structural similarity,computing speed,and image visual perception.The experimental results of the public satellite remote sensing database DOTA,WHU-RS19,UC-Merced-Land-Use,NWPU-RESISC45 and SIRI-WHU show that the PRDNN proposed in this paper discovers fine-grained visual content such as edges and textures in the image,and the performance of PRDNN is better than several of the latest image super-resolution reconstruction algorithms based on deep learning.Experimental results on benchmark test data sets such as Set5,Setl4,BSD100,Urban100 and Manga109 show that,compared with the excellent image super-resolution methods in recent years,the ERSR proposed in this paper has a comprehensive improvement in the accuracy of the reconstruction results and the visual effect,and it still maintains a very small amount of parameters and a very fast reasoning speed.