| Rain is a common weather phenomenon,which will affect the visual effect of outdoor shooting images,fuzzy and overlay image feature information and then affect the performance of computer vision algorithms.In order to improve the visual effect of the image and ensure the stability of computer vision algorithm,image rain removal method has attracted extensive attention and research.In recent years,the research of image rain removal method has made some achievements,but there are still some problems such as the incomplete removal of rain streaks,blurring original information of images and bad real-time performance.In this thesis,the current research progress of image rain removal algorithm is analyzed and the convolutional neural network is applied to the image rain removal algorithm.Firstly,a single image rain removal method based on multi-scale convolutional neural network is proposed.It uses convolutional neural network to extract rain streaks features of rain images and the expansion convolution to make the feature more abundant.The rain removal images are reconstructed by using these features.Experimental results show that the rain removal method based on multi-scale convolution neural network can preserve the original image information and effectively remove the rain streaks in the image simultaneously.For dynamic scenes in video images,the positions of the rain streaks are detected by using the phase congruency feature and the chromaticity feature of the rain streaks firstly and then the single image rain removal algorithm proposed is extended to the rain removal of video image to realize rain streaks removal in space domain.Finally,interframe information is used to realize rain streaks removal in time domain and we can get the reconstructed images.Experimental results show that the proposed method can effectively remove the rain streaks in the video and improve the visual effect of images.Meanwhile the method can maintain the time domain continuity of video. |
PDF Full Text Request