Research On Image Rain Removal Algorithm Based On Fusion Attention Mechanism

The presence of rain can obscure the details of an image,causing a problem of blurring the image background,and seriously affecting the accuracy of outdoor computer vision system tasks.Due to rain blocking image targets,the accuracy of tasks such as target detection and face recognition decreases.Therefore,effectively removing rain streaks and restoring image background have major magnitude and real value.Traditional image deraining methods propose a priori assumption to remove rain by analyzing the statistical characteristics of rain streaks,but the priori information is designed for specific rain streaks,making it difficult to remove complex and diverse rain streaks,resulting in the problem of rain streaks remaining;The image rain removal algorithm based on depth learning realizes rain removal by training a large amount of data to learn the features of rain stripes.However,the method based on depth learning can easily confuse the high-frequency detail components of the rain pattern and the background,resulting in the background being too smooth and losing detailed information.In response to the above issues,this paper,based on the characteristics of rain patterns and backgrounds,combined with attention mechanisms,carried out research on image rain removal algorithms based on fusion attention mechanisms,achieving rain removal and retaining background details.To solve the problem of rain streaks,first,the residual channel prior information is extracted using the maximum and minimum channel differences of image pixels,and then the shallow features of the image are obtained by fusing the information extracted by ordinary convolution;input shallow features into the deep feature extraction network module,and use the dual attention residual module to extract multi-scale rain features for iterative rain removal;The rain free image is restored through the fusion module.Experiments show that PSNR reaches to 38.39,30.22,36.86 d B,and SSIM reaches to 0.982,0.908,0.963 in the dataset Rain100 L,Rain100H,Rain12.Because this algorithm mainly learns to extract the features of rain streaks,without considering the feature extraction of the background at the same time and ignoring the relationship between rain streaks and the background,in order to more effectively restore the details of the background,based on extracting shallow features using ordinary convolution,this paper uses an expanded residual network to extract multiscale features,using expanded convolutions with different expansion rates to expand the receptive field,increase the attention mechanism,and remove multi-scale rain streaks;In the adjacent stage of rain texture removal,LSTM network and GRU network are used to transfer the mapping of rain texture and background respectively,ensuring the removal of multi-directional rain texture and the extraction of background details,and gradually removing rain to obtain a rainless image.ule.Experiments show that PSNR reaches to 38.67,30.45,36.81 d B,and SSIM reaches to 0.983,0.911,0.972 in the dataset Rain100 L,Rain100H,Rain12.