Research On Rain Anh Haze Image Restoration Methods Based On Deep Learning

Image restoration in severe weather situations is one of the basic issues in the field of computer vision,whose target is aiming to obtain clear images by eliminating the degradation of weather.In view of the fact that the widely used image restoration methods are mainly aimed at a single weather condition,this dissertation mainly focuses on the image degradation caused by a variety of bad weather(rain,haze,and attached raindrops)conditions,and uses a single network to study image restoration.The main research and innovation of this dissertation are as follows:First of all,this dissertation researches a method based on attention-guided multibranching for single image deraining and dehazing.This method follows multi-path and multi-branch models,which enables multiple interconnected branches to interact and fuse features at different scales so that shallow feature can guide deeper feature prediction,and introduces a spatial and channel attention module at each interaction node so that it can adaptively learn important information and improve the network expression ability.In addition,a structural detail texture module is introduced to preserve the detail texture information.The results show that this method can not only achieve effectively image deraining and dehazing,but also perform well in image dehazing performance and preserve the background and detailed information of the image completely.Secondly,this dissertation researches a method based on multi-stage progressive for single image deraining and dehazing.This method mainly adopts a multi-stage architecture,which can gradually restore the input degraded images and decompose the whole restoration process into multiple stages.Specifically,this model first learns features of the context using an encoder-decoder architecture and then combines them with the high-resolution image.In the first two stages,a self-attention mechanism,the Vision Transformer encoder basic block,is introduced.At the same time,channel and spatial fusion attention is adopted in the third stage,so that the reorganization unit is used to realize the exchange of important information between features.The results show that the method improves the performance of image dehazing,and can better achieve image dehazing.Finally,this dissertation researches a single image deraining and dehazing method which aims at the phenomenon of rain and haze coexistence in the real world based on decoupled dynamic filters.This method is based on the generative adversarial network framework and uses decoupled dynamic filter modules to replace standard convolutions.Therefore,this improved residual block can be used as the basic block of the network generator,and its generator part is designed with a multi-scale coding structure so that the reconstruction of image can be better achieved.The results show that the network can use a unified model to achieve the purpose of removing rain and haze in the image at the same time.