Research On Deep Image Rain Removal Network Based On Feature Enhancement

With the rapid development of artificial intelligence technology,the requirements for image and video quality in many fields are getting higher and higher.However,in many scenarios,due to the influence of extreme weather(such as rain,snow,fog,etc.),the images directly obtained from outdoor have low contrast and various noise effects,which cannot meet the subsequent advanced computer vision tasks,such as image classification,object detection and recognition,etc.Especially in rainy days,rain noise can block image information,reduce image visibility and contrast,seriously affect image quality,and may cause some vision algorithms to fail.Therefore,it is very necessary to eliminate the rain streak noise in the rainy day image,remove the interference caused by the weather,and restore a clean background image,which is very necessary to improve the performance of subsequent computer vision tasks.In recent years,with the rise of deep learning technology,the performance of image rain removal algorithms has been greatly improved.However,there are still problems such as the loss of image background information caused by excessive rain removal or the residual rain streaks caused by insufficient rain removal.To solve the above problems,this thesis focuses on the single image rain removal method based on deep learning,and proposes two new image rain removal networks.The main research work is as follows:(1)A feature enhanced rain removal network based on attention mechanism is presented.The main network is composed of several joint spatial attention feature enhancement modules in a cascade.The joint spatial attention feature enhancement module consists of a multi-scale large-core attention module and a coordinate attention module.Inspired by Transformer,the multi-scale large-core attention module can increase the perception field for extracting features and enhance the network’s attention to long-distance information features.The coordinate attention module enables the network to focus more on the key locations of the rain features and uses the attention graph to enhance the rain features.The network is used to extract accurate rain streak patterns and to construct residual operations with the rain images to obtain the final rain removal image.A large number of experiments on multiple datasets have proved that this method is superior to some advanced deraining algorithms both objectively and objectively.(2)A feature enhanced rain removal network based on high frequency guide map is presented.Considering that the rain streaks mainly exists in the high frequency component of the image,we use the Butterworth filter to extract the high frequency component as a guide map of the network.First,a spatial feature enhancement module is designed to enhance the initial features of rain images using high-frequency guide maps.Then,the feature enhancement module based on channel attention is used to enhance the structure information of the rain streaks on the channel dimension.Then,in order to extract richer rain patterns,a multiscale feature extraction module is constructed to extract features at different scales by using multibranch convolution with pyramid structure.Finally,the two-way feature enhancement module with two feature enhancement branches is constructed to further enhance the rain streak features to enhance the spatial and structural features.The final rain removal image is obtained by subtracting the rain pattern extracted from the network from the rain image.The network constructs different enhancement modules to extract the rain pattern features,so as to extract the rain layer information more accurately and avoid the problem of inadequate rain removal effectively.Experiments on multiple datasets show that the proposed network has better rain removal performance both subjectively and objectively than some advanced methods.