Research On Image Super-resolution Reconstruction Method Based On Attention And Autoencoder

Image super-resolution is an important technology in the field of image processing,which aims to reconstruct and complete the lost information of low-resolution images by software method,and then reconstruct the clear high-resolution images.In recent years,although the image super-resolution algorithms based on convolutional neural network have been developed vigorously,most of the existing algorithms still simply deepen or widen the network to promote the improvement of network performance,resulting in many deficiencies in the extraction and utilization of image features.To solve the existing problems,this thesis deeply studies the image super-resolution method based on attention and autoencoder.The main research contents are as follows:(1)For the problems that image feature extraction and utilization are not sufficient and important features are not highlighted,this thesis presents a deep-connected multiscale residual attention block,it not only can fully exploit both the scale and hierarchical features,but also can adaptively rescale the channel-wise features to highlight the important features by considering the local interdependencies between channel-wise features.For the problem that existing methods do not utilize the potential correlations between samples of the training dataset,this thesis also designs a deep feature extraction structure.Global-aware external attention module is embedded in the structure to capture the correlations between different samples,which further helps the reconstruction of image super-resolution.At the same time,this structure can also bypass abundant hierarchical features.(2)Existing image super-resolution networks almost reconstruct high resolution images by fitting the direct mapping relationship between low and high resolution images.However,this process is a morbid process of rebuilding lost information from scratch,so it is relatively difficult to obtain good reconstructed image quality.However,self-reconstruction of high-resolution image is a process of sparse coding and decoding of high-resolution image,which is a process from something to something.Therefore,this task is relatively easier than the super-resolution task of image,and the reconstructed image is of better quality than the super-resolution image.Based on this idea,this thesis proposes a single image super-resolution method based on cross-aligned dual-way autoencoder.By designing a cross-aligned autoencoder super-resolution network,it is realized that the self-reconstruction tasks of low-resolution images and high-resolution images are utilized to assist to complete the super-resolution reconstruction task.Through experiments on public datasets,the proposed method based on external attention and multi-scale residual attention can not only fully extract and utilize image features,highlight important channel features,but also take advantage of the potential correlation between different samples to effectively improve the reconstruction performance of the model.In addition,through the analysis of image self-reconstruction task and image super-resolution reconstruction task,a creative single image superresolution method based on cross-aligned dual-way autoencoder is proposed.Through quantitative and qualitative experiments and analysis on multiple testsets,the effectiveness and superiority of the proposed methods are fully proved.This thesis has 29 figures,14 tables,and 88 references.