Research On Image Super-resolution Reconstruction Technology Based On Residual Dense Jumper Network

As a basical research branch in digital image processing,image Super-Resolution(SR)technology has widespreaded into the fields from aerospace imaging,video surveillance to high-definition TV displaying,etc.Therefore,there are universal meanings for SR in aspect of theoretical study and pratical application.Nowadays,SR algorithms could be divided roughly into two categories,that is,reconstruction-based and learning-based method.When the latter solves the inadequate priori information of inputted low resolution(LR)image by introducing large external training data,it is remarkable in restoration performance.The convolutional neural network(CNN)based SR method developed in these recent years pertain to this.However,problems also exist in CNN-based methods.Thanks to the limited shallow depth in constructed model,generative SR images are regularly unimpressive,resulting from the inefficient learned abstract high-frequency features.Forwardly,a deeper network made the computing complexity increased rapidly and solutions is wanting.Therefore,we conduct a study on these difficulties in SR reconstruction.(1)We build a residual dense connected network used in SR process(RDSR)to promote the pleasing quality of SR reconstructed images.In most cases,conventional networks are always in a dilemma,that is,a shallow network(caused by the simplex building blocks)could not learn more abstract high-level features,causing eventually the texture distortions and over smoothness in SR images.Hereafter,very deep(437,to be exact)RDSR model employed the Nested Residual Dense connected Block(RDB)as elementary composition is put forward.For the sake of a much complex mapping function and the broaden receptive fields,much information in LR image are captured.Experiments certifies that RDSR model achieves higher restoration precision in dense texture regions compared with a shallow network(less than 30 layers).As for quantitative assessment,average PSNR values improves 1.22dB(x4)on standard testing datasets when comparing with VDSR.(2)What's more,the pyramid structure based RDSR(pRDSR)is brought forward to reduce the convergence time,and then the complexity of training process.As model become deeper,parameters such as weight,biases increase fiercely,training process has characteristic of difficult trouble-shooting.In image recognition,residual learning strategy could be used to speed up the convergence by learning a sparse information,when laplacian pyramid lower layer is higher sparse,therefore,conventional end-toend mapping is substitute for learning a sparse laplacian residual output,so as to concentrate more computational power on the acquisition of high-frequency component rather than on the remapping of LR plain information.The learning curves in training process prove that pRDSR would converage to a lower value in limited time than RDSR.Furthermore,average PSNR in 5 standard testsets boost 1.13dB(x8)than VDSR algorithm,and the reconstructed textures come more near the original images.