Image Super-resolution Reconstruction Based On Skip Connection Residual Network And Parallel Channels

Resolution is an important indicator of the richness of an image.In the field of image reconstruction,Single-Image Super-Resolution(SISR)indicates the process of reconstructing a single high-resolution image from a low-resolution one.Compared with low-resolution(LR)maps,high-resolution(HR)digital images contain generally more primitive information and therefore reflect better overall visual effects.However,many current SISR algorithms are insufficient to solve with severe degradation of imaging quality in reconstructed image and sometimes even bear high computational complexity.With its superior learning ability,artificial neural network has made the rapid development of artificial intelligence,making artificial neural network once again a research hotspot.At present,deep learning has been widely used in various fields such as computer vision,speech processing,and natural language processing,and has even played a leading role in some fields.The single image super-resolution reconstruction technique aims to reconstruct a low-resolution image through a series of algorithms to reconstruct a corresponding high-resolution image.At present,the more mature methods are based on frequency domain method,non-uniform image interpolation method,convex set projection method,maximum posterior probability method and sparse representation method.This paper focuses on the use of deep learning to achieve single image super-resolution reconstruction.The work carried out in this paper includes the following aspects:(1)Firstly,the research status of image super-resolution reconstruction technology at home and abroad is analyzed,and the main technical methods of image superresolution reconstruction are listed and analyzed.And introduced the basics of deep learning(the basis of artificial neural networks and the basis of convolutional neural networks).(2)Analyze the classic image super-resolution reconstruction technology and propose its own improvement method for its shortcomings.Because the VDSR algorithm uses a one-way propagation convolutional layer,in a very deep network structure,the features in the shallow network cannot be better utilized,so we adopt a residual network of jump connections to achieve fast image detail features.Learn and use the adjustable gradient cropping to achieve fast convergence of the network.(3)The parallel channel network performs a 1×1 nonlinear mapping on image features previously extracted through the deep network,and introduces more nonlinear mappings to the image representation while reducing the total feature map dimensions.Thereafter,the receptive field of the single pixel is further widened by the convolution kernels of 3×3 and 5×5 filters,respectively.Finally,the final high-resolution image reconstruction is done by global residual learning.