Research On Image Super-resolution Reconstruction Technology Based On Multi-scale Convolution Residual Learning

The task of image super-resolution is mainly a technical process of restoring the blurred image generated by the distortion and degradation of an image to a clear and realistic high-resolution image through a series of hardware or software technical methods.This image super-resolution technology has a wide range of application value in the fields of national defense and security,medical image disease analysis and so on.Therefore,the problem of image super-resolution reconstruction has also become a current research hotspot.Due to the large expenditure on hardware,the existing image super-resolution reconstruction methods are mainly widely used on the software level.With the progress of society and industry,the traditional image reconstruction technology is difficult to meet the needs of the development of modern society.With the emergence of artificial intelligence and convolutional neural network,researchers have applied convolutional neural network to the field of image super-resolution reconstruction and achieved excellent results.This technology has quickly become a hot topic in the field of image,but there are also many difficulties.The main difficulties are as follows:(1)because different depth regions of the feature map have different scales of receptive field information,most of the existing networks have a single scale and only extract fixed size local receptive field information,resulting in a single scale of feature extraction.(2)The existing network models do not pay enough attention to the high-frequency texture details of the image,resulting in the edge texture of the reconstructed image is not clear.(3)The image feature information extracted from different depths in the network model can not be transferred effectively,which leads to insufficient feature utilization and affects the reconstruction effect.To solve these problems,this paper designs two image super-resolution reconstruction methods based on SRCNN.To solve problem(1),the paper designs an image super-resolution reconstruction algorithm MCRN based on multi-scale residual network.MCRN connects multiple convolution kernels of different scales in parallel on the network width to form a multi-scale residual unit MRT.The designed multi-scale residual unit fuses the feature information of different scales,which improves the ability of the network model to extract image features of different scales and capture more abundant receptive field information,Combined with the method of jump connection,the information transmission ability between different network paths and network depths on the network width is strengthened,and the feature reuse between different network paths and depths is effectively realized.In order to solve problems(2)and(3),an image super-resolution algorithm IBFN based on multi-level feature fusion network is designed on the basis of MCRN.In order to solve the problem(2),IBFN adopts the designed multi-dimensional attention mechanism MRSF.First,the attention effect on the channel is realized through standard convolution,and then the deep separable convolution DSC is introduced to realize the attention machine effect in the spatial domain,which effectively improves the network’s attention to the high-frequency information in the channel and spatial domain,allocates more training resources to the high-frequency information,and effectively reduces the parameters.In order to solve problem(3),IBFN designed a multi-level feature fusion connection(MNL).Through the multi-level jump connection,the output features of the constructed residual blocks can be effectively transmitted,and the utilization of image feature information at each depth can be improved.At the same time,the output features of each residual block are fused to obtain rich feature information,Finally,in the reconstruction phase,the compensated reconstruction module(CR)is used to repair the upsampled image and reconstruct the high-resolution image.The paper takes DIV2 K data set as the training data set to train the MCRN model and IBFN respectively.Two numerical evaluation indexes,peak signal-to-noise ratio index(PSNR)and image structure similarity index(SSIM),are used to objectively evaluate the image quality generated by the model,and then the subjective visual characteristics of the generated image quality are identified and analyzed according to the human visual field,The quality of the super-resolution reconstructed image is evaluated.The methods in this paper and several advanced methods are applied to the test data sets Set5,Set14 and BSD100 × 2,× 3,and × 4 scale factor reconstruction operation.Through the experimental data,it is concluded that the data indexes PSNR and SSIM of the algorithm in this paper are higher than those of other advanced methods.In terms of visual effect,the high-resolution images generated by MCRN and IBFN have richer texture details and higher definition,and the overall effect is closer to the original high-resolution images.