Image Super-Resolution Reconstruction Based On Deep Residual Network

Nowadays,high-resolution images play an important role in medical,military and social et al security fields,how to acquire high-resolution images under limited conditions has become a popular research.Due to the cost and technical limitations of hardware imaging equipment,reconstruction of high-resolution images by software is an important implementation method,but many reconstruction algorithms suffer from incomplete image feature extraction and computational difficulties in the training process.In order to further improve the image reconstruction capability,this paper investigates the image super-resolution algorithm based on residual network and proposes two improved algorithms,the main work includes the following two aspects.(1)To address the problem of incomplete mapping of deep features and poor reconstruction due to easy loss of mapped features in the residual network algorithm,an image super-resolution reconstruction algorithm(Multichannel Deep Residual Network,MDRN)based on multi-branch depth residual network is proposed.The algorithm introduces a multi-branch architecture in the deep feature mapping part and introduces a residual structure in each branch to increase the network width while ensuring the network depth,so that the network model can map multiple types of information;and improves the residual structure in each channel to reduce the memory resources required by the algorithm;then the information mapped by multiple branches is superimposed in the deep feature mapping output layer to further make The information of multiple branch mappings is then superimposed in the deep feature mapping output layer to further make feature fusion and facilitate information transfer.(2)To address the problem of increasing model processing difficulty due to the increase of residual network depth and width,an image super-resolution reconstruction algorithm(Residual Attention Neural Network,RANN)with fusion of residual attention mechanisms is proposed.The algorithm first uses a convolutional layer to extract the underlying features;then introduces a fused attention mechanism in the residual block of the deep feature mapping part,assigning different weights to the features of different depths from both channel and space dimensions,focusing the residual learning on more important features such as image details,and maximizing the network performance with limited resources;and introduces a jump connection in the output part,overlaying the output of each residual The output of the attention module is superimposed to ensure the comprehensiveness of the extracted features and avoid information loss.The algorithms in this paper are tested on three publicly available datasets,Set5,Set14 and Urban100,and compared with the reconstruction results of various algorithms such as Bicubic,SRCNN and VDSR.The results show that the two algorithms proposed in this paper have better performance and can reconstruct clearer images with better visual effects and objective quality evaluation.