Research On Video Rain Streak Removal Using Prior Information Tensor Model

With superior real-time and storage performance,outdoor computer vision systems have high application value in the fields of traffic,public security,and identification detection.However,affected by environmental factors such as outdoor rainfall,its captured video images have the problem of occlusion or missing,which is not conducive to the processing and application of post-level systems.For this reason,this paper carries out research on video image de-rain algorithm,and the main work is as follows:(1)A video rain removal tensor model based on total variation regularization with L_2,1norm is proposed for the problem of rain streaks occlusion in rainy videos.The method firstly preprocesses priori information of the rain streaks components and video background to obtain the basis for the construction of the corresponding regularization conditions,which facilitates the rain streaks separation.Secondly,on this basis,considering the existence of irregular dynamic objects in video,the total variation regularization term is introduced to suppress the background intensity variation and mitigate the rain streaks misclassification.Alternating direction method of multipliers(ADMM)is used to solve the proposed tensor model.The experimental results show that the method can effectively remove rain streaks while retaining more background detail information in the presence of dynamic background,and the results are better in three comprehensive quantization indexes,namely peak signal-to-noise ratio,structural similarity and residuals.(2)In order to better solve the problem of rainy video restoration with dynamic objects,a video rain removal tensor model based on total variation regularization low-rank decomposition is proposed.In consideration of the influence of moving objects on the low-rank structure of video background,the rainy video is further decomposed,and the static background is firstly extracted based on the preprocessing of the a priori information of the video with rain.Secondly,the alignment change of dynamic objects is obtained by block matching method to obtain the low-rank property,which is extracted by the low-rank decomposition model,and then the de-rain is completed.Alternate direction method of multiplier(ADMM)is used to complete the model solving work.This method can effectively remove rain lines while retaining more video detail information,and can achieve a better comprehensive quantization index.