The Study And Application Of Raindrop Removal For Single Image Based On Deep Learning Method

Rain removal is an important research field for the application of computer vision.The captured image will be deformed and blurred due to rain,which will affect the performance of the relevant image processing algorithms.In recent years,autonomous driving technology has developed rapidly.Machine learning algorithms based on camera and millimeter wave radar have been able to handle the problem of autonomous driving in a simple situation.However,this technical solution has higher requirements for the driving environment.The occlusion will greatly reduce the accuracy of the autonomous driving system,so rain removal research is of great significance for the application of autonomous driving.This thesis uses the GAN network to perform raindrop removal for a single raindrop image.The main research contents are as follows:(1)Based on the previous research of the single image raindrop removal by using the GAN network,this thesis modified the original network for the autonomous driving application.To improve the computing speed of the network,we reduced the number of network layers,ensuring that the raindrop removal effect can significantly reduce while still having a relative fast computing speed,as far as possible to achieve the high real-time requirements of the autonomous driving system.(2)To further verify the accuracy of the raindrop removal network after the improvement,we input the output of the raindrop removal network into the multi-object detector.For the real-time and accuracy requirements of the target detector for the autonomous driving application,we choose YOLOv3 as the test object detector.We input the raindrop removal image which is output from our network and the corresponding raindrop image to the object detector,the result shows that the object detector can correctly detect the object in the raindrop removal image compared with the corresponding raindrop image,which shows the feasibility of applying our GAN based raindrop removal network in an autonomous driving system.