Research And Implementation Of Image Rain Removal Method Based On Recurrent Generative Adversarial Network

Rain is a very common weather in real life,which not only affects human vision,but also seriously affects the performance of computer systems,including video surveillance,object detection,and autonomous driving.With the improvement of computing performance and the development of deep learning theory,the image de-raining task has made many breakthrough advances.However,there are also problems such as blurred background and residual rain marks.This article summarizes the advantages and disadvantages of the existing image rain removal methods and their application scope.Aiming at the problem of image rain removal in real scenes,this paper focuses on the characteristics of rain images,and proposes an image rain removal method based on decomposed cycle generative adversarial network.It conducts in-depth and systematic research on training using unpaired data.In addition,the characteristics of rain images in natural scenes are studied in depth,and an image rain removal method combining self-supervised constraint and cycle generative adversarial network is proposed to improve the performance of rain removal algorithms in practical applications.The specific work of this article is as follows:First,this paper proposes an image de-raining method based on decomposition of a recurrent generation adversarial network.The cycle generative adversarial network(Cycle GAN)model can only be mapped one-to-one in the image conversion task,and cannot solve the problem of one-to-many mapping in the image deraining task.This paper introduces a rain mask to better generate a variety of rain streak,so that it can use unpaired data training strategies in the image deraining task;a rain streak image generator is added to decompose the input rain image into a background image and rain streak images,so that these two features can be learned at the same time,and the corresponding loss function and learning process are given,which improves the performance of the image de-raining algorithm using unpaired data training mechanism in the synthetic data set.Secondly,this paper proposes an image raining method based on cycle generative adversarial network and self-supervised constraints.Based on the decomposed cycle generative adversarial network,a background guidance module and rain mark guidance module for natural rain images are proposed,and corresponding loss functions are given.Through these two self-supervised constraints,the effectiveness of the unpaired data training mechanism on the realistic rain image dataset is verified,and the rain removal performance of the rain image in the real scene is improved.Finally,based on the method proposed in this paper,an image rain removal system is designed and implemented.This system can not only observe the rain removal effect on a given image,but also show the algorithm execution time.Through the design of front-end display and back-end algorithm separation,the system can also show the effects of other image rain removal methods.