Research On Deep Network Structure For Image Restoration

Image restoration is to process the missing or damaged parts of the image data through a complex algorithm of the computer,so as to reconstruct or restore the degraded image.In recent years,deep neural networks have been extensively studied in the field of image restoration due to their powerful ability to autonomously learn features;Many researchers at home and abroad have invested energy and time to study the design of their network structure,and achieved remarkable image restoration effects.This paper aims at two major themes of image restoration:image super-resolution reconstruction.and image denoising,and proposes the corresponding network structure design.The results achieved are as follows:An image hyper-reconstruction network based on multi-scale distributed network(MSN)is proposed.Firstly,use different scale convolution kernels to capture multi-scale features of low-resolution images;secondly,feature maps captured by convolution kernels of the same size are directly input to their corresponding multi-scale hy brid group(MHG)for feature training and learning;all the multi-scale hybrid group(MHG)training feature maps are cascaded to obtain small-size feature images;finally,Meta upsampling is used as an image enlargement module to amplify the trained feature images with any scale factor to obtain image super-resolution reconstruction.Among them,the mixed convolutional layer of each MHG is composed of hole convolution and standard convolution.The hybrid convolutional layer can fully learn higher-level details from the previous and current scale convolutional layers,and the output of each hybrid convolutional layer is fed back to the subsequent hybrid convolutional layer through a jump connection,thereby generating dense connections to form a very deep and effective super-resolution reconstruction network MSN.Through experiments and evaluation of multiple images,it is found that the MSN network proposed in this paper is superior to the current advanced methods in numerical results and visual quality.A dense network for image denoising based on separate hybrid convolution is proposed.The dense network is composed of two symmetric modules:a separation module(SNM)and an error feedback module(EFM);the feature learning convolutional layers of the two modules all adopt a hybrid convolution design,which can effectively capture the small features of the image and smooth the image noise;the mechanisms of residual learning and jump connection between the two hybrid convolutional layers used in the two modules make the network dense and efficient.Among them,the difference between the noise image and the clean image is firstly captured and trained by the noise separation module(SNM);the error feedback module(EFM)performs image detail recovery and further removes the image noise on the pre-cleaned image obtained by training.Through experiments and evaluation of multiple images,it is shown that the dense network proposed in this paper can not only remove noise,but also retain clear image details;compared with the current advanced denoising method,the denoising performance of this network is outstanding.