Research On Single Image Deraining Based On Residual Network

Nowadays,automatic driving and video monitoring technology are widely used.However,there are many difficulties in collecting pictures or videos in rainy days.The rain stripes will seriously block the image background,resulting in the degradation of the visual quality of the collected images and videos.In rainy days,the visibility is low and the background scene is blocked.A series of features such as contrast and color of the target in the image will be relatively weakened,resulting in the unclear description of the target image,making the video or image system can not work normally.The information of the rain water layer and the background layer in the image are complex coupled.The high-frequency details of the original information can not be retained in the image after removing the rain layer.The relevant feature information is lost.Whether the rain can be removed from a single image and detail information in the image can be better recovered is an important issue.In view of the above problems encountered in the task of removing rain,this paper constructs a real rain data set and uses deep learning algorithm to extract features,so as to bring better rain removal effect for the real rain image.The main work is as follows:First of all,the cyclic residual network is used to derain by iteration.The results of each iteration and the original rain image are taken as the input of next iteration.The residual learning is used to predict the rain stripe layer,and the rain removal problem is solved in multiple stages.This method significantly reduces the network parameters,while the performance does not degrade.The loss function is optimized and the multi objective loss method is more suitable for our network.The method finally achieves PSNR of 31.80 on synthetic data set and SSIM reaches 0.93.The experiment shows that this method can remove the rain layer of the synthetic rain data set more effectively,but it is not good on the real rain image.At present,the synthetic data set is used to train the deraining network.There is deviation between the synthetic image and the real image,so we construct the real rain image data set.The construction method is to shoot rain and clean pictures in the same position and angle,including different rainfall degree and background.In order to focus the algorithm more on the location of rain and enhance the effect of the output image,the attention module based on two rounds and four directions is introduced.The first round of IRNN aims to generate the feature map of the adjacent context of each position in the input image,and the second round of IRNN further collects the global information.The improved method is compared with JORDEN and RESCAN methods in the real rain data set.Our method achieved good results which PSNR reached 26.95 and SSIM reached 0.83.