Rain And Snow Removal Algorithm In Single Image Based On Sparse Coding

Computer vision system is widely used in target detection,feature extraction,object recognition and many other areas,but bad weather,such as rain and snow,will degrade the quality of video and image.It severely affects the performance of computer vision algorithms and leads to a significant reduction in the effect.As a result,rain and snow removal in the images helps to improve the effectiveness and accuracy of various types of computer vision algorithm.In general,depending on the type of data to be processed,the research in rain removal includes two kinds,the algorithm based on video data and the algorithm based on single image.We can utilize the information of more frames in the algorithm based on video.This area has significant progress while the algorithm based on single image is more difficult,for the reason that the available information is less.In this paper,the existing video and image rain removal algorithms are studied and compared.As for the algorithm in single image,the paper puts forward two new algorithms,which improve the effect of rain and snow removal in single image.As for the existing classic algorithms based on image decomposition,this paper proposes a new method to optimize geometry component in high frequency part.First of all,the smoothing filter is applied to decompose the image and the high frequency part of rain image is got.Then we combine sparse representation and affinity propagation to obtain the rain component in high frequency part,and the high frequency part minus the rain component,followed by a smoothing filtering.The result is the new geometry component.In addition,the dictionary got from sparse representation is classified again,which optimize the distinction between rain component and geometry component.Finally,we complete the image restoration.In addition,for the reason that the current rain removal algorithm in single image based on image composition has shortcomings,the effect of these algorithms is heavily dependent on sparse representation and dictionary learning.So this paper proposes an algorithm based on low rank representation.First,bilateral filtering is used on rain image to get the low frequency part and high frequency part,and we split block and calculate HOG feature in high frequency part.All of the HOG feature vectors are merged into a feature matrix as the input of low rank representation.Then,the coefficient matrix got from low rank representation is clustered,and through the value of variance,we select the rain sub-blocks to be removed.Finally,we add the result of the last step to low frequency part,and the restoration of non-rain image is finished.Experiments prove that the two algorithms proposed in this paper can keep more details of image,and the rain and snow removal is better.