Research On Removing Rain From Single Image

Life is often accompanied by bad weather.Rainy days will greatly reduce the quality of captured images and videos,affect people's observation and the performance of computer vision system for detection and recognition.Therefore,it has important application value to study the rain removal algorithm of images.According to the type of data processed,there are two types of the rain removal algorithm research:video rain removal and single image rain removal.The key to rain removal algorithm is to make the rain clean,while minimizing the loss of other details.This thesis systematically summarizes the algorithms of video and single image rain removal,and mainly studies the problem of single image rain removal.The main work is as follows:(1)Based on image decomposition,a rain removal algorithm is studied which can accurately extract raindrops and preserve useful image details.The high-frequency image directly obtained by image decomposition contains raindrops and effective image details,removes the detail information by using the low-frequency image to generate a detail preserving template.Using the low-frequency image as the guide image to guide and filter the high-frequency image and decomposes the high-frequency image again,thereby suppressing the detail information,and obtaining a high-frequency image with more raindrops and less background details.Perform threshold segmentation on the newly obtained high-frequency image to obtain a template for extracting rain marks,and use this template to multiply the original high-frequency image to obtain the raindrops image.Finally,subtract the raindrops image from the original image to get the rain removal result.(2)Raindrops have obvious directional consistency.According to the directional correspondence between the spatial domain and the frequency domain,studies a rain removal algorithm based on directional filtering in frequency domain.When transforming the image into YCbCr space,mainly distributes raindrops in the luminance component and the sensitivity of human eyes to luminance is higher than that of chrominance component,so adopting different rain removal treatments for luminance component Y and chrominance component CbCr.For the luminance component,estimates rain mark direction by edge detection and gradient direction statistics and designs a fflter to suppress the high-frequency components in the Fourier spectrum perpendicular to the spatial raindrops to achieve rain removal while maintaining effective image information.Using guided filtering on the original rainy image and converting it to YCbCr space,the chromaticity component that suppresses rain marks will be obtained,and obtains the final rain removal result by color space conversion combined with the processed luminance components.In this thesis,experiments are carried out on synthetic and real rainy images,comparing with other related rain removal algorithms.Subjective visual evaluation and objective indicators to measure the similarity between rain removal results and original images verify the effectiveness of the proposed algorithm.