Image Restoration Algorithm Research Based On Convolutional Neural Network

Digital image is one of the most basic data used in people's daily life.At the same time,the quality of digital image also seriously affects the accuracy of other visual tasks.The image restoration algorithm aims to alleviate this problem by reconstructing corresponding high-quality images from low-resolution and blurred images.In-depth research on image restoration algorithm based on convolutional neural network,and propose solutions to existing problems.The specific summary of research work and innovations are as follows::First,the image restoration algorithm based on deep learning is summarized,compared and problem analyze.Aiming at the problem definition of single image super-resolution and single image dehazing algorithms,the classic structure and basic modules are summarized and compared,and their characteristics and deficiencies are analyzed in-depth.The training strategy of the algorithm model adopted is explained in detail.The experiment compares the respective advantages and performance of the image restoration algorithm,and points out the existing problems,providing motivation and theoretical background for the innovation of the algorithm in the subsequent chapters.Second,a single image super-resolution algorithm based on cascading feature pyramids is proposed.Aiming at the problem of insufficient utilization of existing algorithms in multi-level feature information,a cascading feature pyramid module is designed to expand the multi-scale characterization range of features and eliminate the semantic gap between multi-level features through convolution cascade and layer-by-layer fusion structure Enhance the multi-level feature fusion effect.In addition,an improved asymmetric residual Non-Local module is proposed,which uses a lightweight self-attention mechanism to efficiently aggregate global context information.The PSNR index of the 4x super-resolution result of the algorithm on the Set5 data set reaches 32.66 db,and it has good generalization under various degradation models.Third,a single image dehazing algorithm based on progressive local connections is proposed.Design a lightweight progressive feature fusion module to improve network learning efficiency,which consists of two parts: progressive local connections and feature attention.Specifically,the progressive local connections extracts multi-layer features for fusion by layer-by-layer separation and local stitching.Feature attention uses channel and spatial dual attention modules to evaluate the importance of features.In order to efficiently aggregate the global context information,an improved non-symmetric residual Non-Local module is proposed to improve the image dehazing effect while maintaining its light weight advantages.The PSNR value of the evaluation index of the algorithm on the SOTS indoor test set is increased from the best 36.39 db to 38.02 db,and at the same time save a lot of model parameters.