Research On Image Super-resolution Algorithm Based On Residual Channel Attention And Generative Adversarial Network

Image super-resolution is one of the important methods to improve image resolution,and image super-resolution techniques are widely used in various computer vision tasks.Since convolutional neural networks have a strong feature representation capability,the introduction of convolutional neural network methods in image super-resolution can significantly improve the reconstruction quality of images.It can be seen from the comparison between different superresolution models that the network performance gradually improves with the depth of the model.However,there are some limitations of these methods.First,deeper and wider networks mean at the same time a larger computational effort,which can increase the training difficulty of the network.Second,many models treat the information of each channel equally,and channels with more useful information do not get more attention,and the network lacks flexibility.In addition,some reconstructed images suffer from the problem of losing high-frequency edge information.To address the above problems,this thesis makes full use of the correlation between channels,combines channel attention,generative adversarial networks,and edge enhancement to propose a convolutional neural network-based image super-resolution algorithm,and conducts experiments to test the performance of the proposed algorithm.The research content of this thesis is as follows.(1)An image super-resolution algorithm based on multi-scale channels and pixel attention networks is proposed.In this thesis,we use end-to-end convolutional neural networks to construct an image super-resolution model,and propose an adaptive structure for assigning features by exploiting the interdependencies between each channel as well as pixels: a multiscale channel and pixel attention network,which enables the network to be more flexible in processing different types of information,allowing the extraction of more and more useful information,and solving the problem that some methods based on deep learning are the problem of lack of flexibility when processing different types of information.In addition,introducing a pre-processing block before deep feature extraction enhances the network’s ability to adapt to image features at different scales,effectively improving the reconstruction quality of the model.The method can notice more useful information and thus infer finer image details.(2)A multi-level wavelet edge-enhanced image super-resolution algorithm based on generative adversarial networks is proposed.The method constructs an image super-resolution network model by generating adversarial networks,and proposes an improved multi-level wavelet channel and pixel attention network by combining multi-scale channel and pixel attention network with discrete wavelet transform,which can increase the perceptual field while taking into account the computational cost,and helps generate intermediate SR images with richer texture details.In addition,the method introduces an edge enhancement module to enhance the edges in the intermediate SR images,which solves the problem that some methods of expanding the perceptual field tend to cause heavy computational costs and lose high-frequency edge information in the reconstructed images.By using this method,the network can reconstruct the real image more accurately and generate a final SR image with clear edges and no noise contamination,which effectively improves the image quality.