Research On Image Deraining Algorithm Based On Attention Mechanism And Feature Fusio

Outdoor vision systems require image acquisition in real-world environments,however,shooting in adverse weather conditions can lead to reduced image quality and seriously affect the performance of the vision system.Rain is one of the common inclement weather conditions,and rain can cause blurred images or image backgrounds to be obscured,causing significant problems for subsequent image analysis and processing.Image deraining is therefore a challenging image pre-processing task.Convolutional neural networks,with their powerful feature representation and learning capabilities,have achieved good results in the field of image rain removal.However,existing rain removal methods still have problems in terms of rain removal performance and model complexity.To address the shortcomings of existing methods,this thesis combines existing deep learning techniques to conduct in-depth research on image rain removal,and the main work and results achieved are as follows.(1)In order to effectively remove rain streaks from images and retain a large amount of background detail information,this thesis proposes a multi-scale feature fusion image deraining method based on an attention mechanism.The feature extraction stage consists of multiple residual groups containing two multi-scale attention residual blocks.The multi-scale attention residual blocks use the multi-scale feature extraction module to extract as well as aggregate feature information at different scales and further improve the feature extraction capability of the network through coordinate attention.Local feature fusion is performed within groups,and the global feature fusion attention module is used between groups to better fuse features at different levels and to focus the network on rain pattern regions through pixel attention.Compared with other existing image deraining algorithms on simulated and real rain image datasets,the quantitative and qualitative metrics of the proposed method have significantly improved,and the visual effect of the deraining image is better.(2)In order to reduce the complexity of the model,reduce the network parameters and make the network lighter while obtaining better rain removal capability,this thesis proposes an image rain removal algorithm based on the distillation of residual features by an attention mechanism.The method gradually removes rain streaks from the network by a cyclic approach.Within each stage,a multi-channel split concat block is first used to extract and aggregate the rich rain streak features,while a long and short-term memory network is introduced to guide the next stage of rain removal.Then,feature distillation is performed using a lightweight residual feature distillation network to refine the features,and finally,a dual attention block is applied to the distilled features to adaptively focus on the important feature information.Experiments show that the proposed method can obtain deraining images that are closer to the original image with a small number of parameters and a simple network structure.