Research On Key Technologies Of Image Super-resolution Reconstruction In Wavelet Domain

Image is a fundamental medium for humans to perceive the external world.With the advancement of technology,there is an increasing demand for higher image quality.Image resolution is often limited by factors such as acquisition devices,compression techniques,and imaging technologies.Image super-resolution aims to reconstruct high-resolution images with rich details from low-resolution ones,which is not only of significant research value but also a highly challenging problem.In recent years,super-resolution methods based on deep neural networks have become mainstream and widely applied in various fields such as intelligent transportation,medical imaging,aerospace,communication,and entertainment.However,traditional deep learning methods in spatial domain often struggle to effectively extract detailed and structural information.On the other hand,wavelet analysis offers the advantage of decomposing signals into different resolutions,allowing independent processing while gradually sampling high-frequency signals to focus on object details.This insight inspires super-resolution reconstruction techniques.Therefore,the dissertation focuses on the key technologies of super-resolution reconstruction with deep neural networks and wavelet theory,particularly in enhancing image quality and model lightweighting.The main research areas include the following aspects:1)To address the problem of treating all subbands in the wavelet domain in the same manner,a super-resolution reconstruction network based on wavelet frequency separation attention is proposed.This network separates the high and low-frequency subbands of the transformed low-resolution image and designs an additional branch network dedicated to learning high-frequency features.In the feature extraction network,ghost extension blocks are introduced to reduce the parameters required for feature acquisition.Moreover,an improved attention ghost extension block is proposed to enhance the quality of highfrequency feature extraction.Finally,the high and low-frequency subbands generated by the two branch networks are reconstructed into a high-resolution image through inverse wavelet transformation.Experimental results show that the model maintains lightweight while achieving improvements in subjective visual and objective metrics.2)To address the problem of single-resolution structures reconstructing only one scale of wavelet coefficients at a time,a super-resolution reconstruction network based on deep wavelet Laplacian pyramids is proposed.This network combines the Laplacian pyramid structure to simultaneously predict multiple scales of wavelet coefficients.Meanwhile,different feature extraction module designs were explored to adapt to the learning of wavelet coefficients.In addition,a texture-charbonnier loss is designed to reconstruct the wavelet coefficients,thereby learning both low and high-frequency information.Compared to other progressive reconstruction models,this work achieves overall better results and demonstrates improved robustness in the face of degradation such as Gaussian noise and motion blur.3)To address common issues in progressive super-resolution models,such as the one-sided consideration of subspaces,and sole focus on either wavelet or spatial domain information,a super-resolution reconstruction network based on wavelet multi-resolution transformation analysis is proposed.This network captures information from multiple subspaces through a multi-resolution structure and perceives the interdependency between wavelet domain and spatial domain features by adaptive fusion modules.Additionally,a convolutional filtering-based wavelet transformation module is designed to support wavelet decomposition and reconstruction in the network,serving as a fundamental module that requires no training.This module achieves advanced levels of accuracy and runtime efficiency.The reconstruction results excel in both objective and subjective quality aspects.4)To address issues of singular transmission of information in multi-resolution structures,and the lack of attention to signal distribution texture in wavelet coefficient prediction,a wavelet pyramid recursive super-resolution reconstruction network based on wavelet energy entropy constraint is proposed.This network transmits previous-level wavelet coefficients and additional shallow coefficient features and considers parameter sharing between high-and low-frequency coefficients within and across pyramid levels.A multi-resolution wavelet pyramid fusion is devised to promote information transmission within and across resolution levels.Additionally,a wavelet energy entropy loss is proposed to constrain the reconstruction of wavelet coefficients from the perspective of signal energy distribution.Experimental results on public datasets demonstrate that the proposed network not only enhances model efficiency but also maintains superior performance in both quantitative and qualitative evaluations.