Research On Image Super-resolution Based On Attention Mechanism And Multi-scale

Image super-resolution reconstruction is a hot research direction in the field of computer digital image processing.In practice,this technology has been widely used in many fields such as military,medicine,surveillance,and remote sensing.In recent years,deep learning has been widely used in the field of image processing,and a major breakthrough has been made in the research of image super-resolution reconstruction technology.However,the current image super-resolution reconstruction technology based on deep learning still has some shortcomings,First,there is a lack of discrimination between high-frequency detail information and low-frequency global information in the image,and high-frequency features cannot be used efficiently,resulting in a lack of texture details in the reconstructed image;Second,only the feature information of a single scale can be extracted,which will result in the loss of some key information in the reconstructed image.This article will focus on the above two points and propose two image super-resolution reconstruction methods to solve the above problems.(1)Propose an image super-resolution model based on attention mechanism and multi-scale feature fusion.This method uses Generative Adversarial Network combined with a channel attention to establish feature dependence between different channels,and adaptively learns to adjust the weights between different frequency features,so that the network model is trained to focus more on high-frequency features.Combined with the parallel generator network model of multi-scale feature fusion,the characteristics of different network layers can be extracted and the feature information of different scales can be merged,so that the generator network model can learn more complete information on the feature map.At the same time,the discriminator after adding the channel attention is more sensitive to high-frequency detail features,and more strictly supervises the generation of the network.The experimental results show that the image quality reconstructed by this method is greatly improved.(2)Propose a residual structure combining Octave Convolution and channel attention,and then combine with SRGAN to propose a multi-scale image super-resolution reconstruction model based on Octave Convolution and attention mechanism.Octave Convolution can divide the feature map into a high-frequency part and a low-frequency part,first separate and then perform multi-scale feature fusion.In the separation process,by compressing low-frequency features,reducing low-frequency redundancy,while increasing the receptive field in disguised form,more contextual information can be obtained.In order to further improve the capture of high-frequency features,channel attention is added to the separated high-frequency features and low-frequency features to make the attention on the high-frequency and low-frequency parts more concentrated,and then the processed by Octave Convolution High-frequency and low-frequency features are multi-scale feature fusion to enrich the content of the generated image.The generator model with improved residual structure can focus on high-frequency features while reducing the influence of low-frequency features,thereby more accurately extracting and fusing features of different scales,and generating images with rich details.The experimental results show that this method can effectively improve the effect of image super-resolution reconstruction,and it also shows that the attention mechanism and multi-scale feature fusion can improve the performance of the image super-resolution reconstruction model.