Image Super-Resolution Reconstruction Research Based On Deep Learning And Attention Mechanism

High-resolution images are widely used in many scenes.However,in the process of acquiring real images,the limitations of the imaging environment and imaging hardware and various noise interferences make it difficult to obtain high-definition images.People often use signal processing and image processing methods to recover high-resolution images from existing low-resolution images through software algorithms.With the emergence of deep learning,the image super-resolution method based on deep learning has become a hot research topic at home and abroad.The attention mechanism allows deep learning models to capture and strengthen local key information to suppress non-significant features,thereby making the network output features more distinguishable and improving the overall performance of the network in the field of machine vision.The SR problem is a low-order machine vision task.Currently,most deep learning-based image super-resolution reconstruction methods do not distinguish the characteristics of low-resolution images.During the training process,the image is given the same weight in different channels and different positions,which leads to the enlarged image is easy to blur and appear artifacts in the place where the texture is rich in details.In order to make the network better reconstruct the details of the image and improve the quality of the reconstructed image,three different convolutional neural networks are combined with attention mechanism to study the image super-resolution reconstruction method.(1)Firstly,the research status of image SR reconstruction methods in detail is summarized and four types of image SR methods based on deep learning are analyzed.On this basis,with the global residual structure as the reference network,an SR algorithm with attention mechanism is proposed.All SR models studied in this paper learn to restore high-frequency details of images in low-resolution space,thereby improving the speed of image reconstruction.For multi-scale SR reconstruction,this paper uses the trained x2 SR model to initialize the multi-scale SR network,which not only speeds up the model reconstruction speed but also improves the SR reconstruction image quality.(2)In order to make full use of the high-frequency information of the image,five different fusion strategies are proposed to obtain the hybrid attention mechanism module,and combines the hybrid attention mechanism module with three different network structures to effectively improve SR reconstruction image quality.(3)Add the attention mechanism module to the residual block to form the RCSA network,and verify the effectiveness of the combination of the attention mechanism module and the residual module on the image super-resolution reconstruction task;comparative analysis uses mixed attention under the fusion of different strategies The force module improves the quality of the residual network reconstruction image.The experimental results show that the use of the tandem fusion strategy to add the attention mechanism module to each residual block can best improve the network effect.(4)In order to strengthen the information flow of the network,obtain high-frequency features at different levels to reconstruct low-resolution images.In the third chapter,the DCSA network is composed of dense modules that add attention mechanism.In a densely connected network,too many attention mechanism modules cascading and stacking will cause information loss and damage network performance.Choosing the maximum fusion strategy and adding the attention mechanism module to some dense blocks enables the DCSA network to fully utilize the advantages of dense connection and attention mechanism,and improve the PSNR and SSIM values of the reconstructed image.(5)In order to reduce the loss of information caused by the superposition of attention mechanism modules,the use of residual connections can reuse features and dense connections can explore the advantages of new features.In Chapter 4,this paper proposes a residual-intensive module with attention mechanism,which fully verifies the performance improvement of RDCSA network brought by local residual and attention mechanism.