| Image resolution is an important index to evaluate the quality of digital image.People always expect the image to have higher resolution.However,affected by many factors such as environmental noise,imaging equipment accuracy and transmission bandwidth limitations,people often obtain images with low resolution,which can not meet the application requirements.Image super-resolution reconstruction refers to the process of using digital image processing technology to restore a high-resolution clear image from one or more lowresolution degraded images.It can provide more image details and improve image quality while improving image resolution,so it has broad application prospects.In recent years,image super-resolution reconstruction based on convolution neural network has been widely studied.On the basis of studying the existing image super-resolution methods,this thesis proposes a new residual fusion convolution neural network model for single image super-resolution reconstruction,and discusses its optimization method,aiming at the problems of insufficient utilization of residual features,loss of details and difficulty in training deep network in the existing methods.The main contributions of the thesis are as follows:(1)A super-resolution reconstruction network based on residual fusion dual attention(RFDAN)is proposed.The network model forms a two-layer residual fusion structure through residual nesting and skip connection,so as to preserve and fuse the residual features extracted from each layer of network,and enhance the network’s high-frequency feature expression ability.At the same time,the model introduces a lightweight space-channel dual attention module to enhance the discriminative learning ability of the network.The combination of two-layer residual fusion structure and dual attention mechanism enables RFDAN in this thesis to achieve a trainable deep network and acquire strong residual feature learning ability.Two times,three times and four times of super-resolution reconstruction contrast experiments and ablation experiments were carried out on four data sets Set5,Set14,BSD100 and Urban100.The experiments proved the effectiveness of the RFDAN model in this thesis and the excellent image detail reconstruction ability.(2)From the perspective of network receptive field,the RFDAN model in this thesis is optimized,and the residual fusion super-resolution reconstruction optimization network based on multi-scale receptive field expansion is proposed to further improve the reconstruction accuracy.Considering the influence of receptive field on attention network,a multi-scale receptive field expansion module is designed to enrich the receptive field of attention network and combine it with RFDAN.The receptive field expansion module uses three sizes of expansion convolution cores,so that the residual features of a larger range and different scales can be captured and effectively fused.The contrast experiment of the super-resolution reconstruction algorithm before and after optimization and the existing mainstream methods shows that the receptive field expansion module can improve the performance of the RFDAN model in this thesis and further improve the quality of the reconstructed image.Moreover,the method in this thesis is superior to the existing mainstream methods in all four data sets,both in terms of subjective visual effects and objective evaluation indicators. |
