Raindrop Removal From Image With Generative Adversarial Network

In the research process of image processing,whether the captured image is clear is important for analyzing the characteristics of the image.Among the various factors that affect the clarity of the image,bad weather has a serious impact on the image capture,and rainy weather is inevitable.One of the factors.Therefore,the removal of raindrops and rain streaks on images taken on rainy days is an indispensable research process to ensure image clarity.At present,the methods for removing raindrops on attached images mainly include traditional methods,based on non-generative countermeasure networks and methods based on generative countermeasure networks.Among them,the traditional method is time-consuming and laborious,and the effect is not strong.The non-generative countermeasure network method for rain removal is still lacking in restoring the image clarity.The generalization needs to be enhanced.The existing methods based on the generative countermeasure network to remove raindrops are not Too much attention to the rain area leads to the problem of ignoring the global dependence of the image,and the removal of rain streaks is missing the feature extraction of rain streaks.In response to the above problems,this paper proposes an improved generative confrontation network to study the removal of raindrops and rain streaks attached to pictures taken on rainy days.(1)Aiming at the problem of raindrops attached to images,this paper proposes a method that combines self-attention and generative confrontation network.Due to insufficient attention to the global dependence of the image,a self-attention layer is added to the self-encoding structure of the generation network to obtain the global information of the image;in order to improve the clarity of the image from the detail level,a multi-scale discriminator is introduced to distinguish between different images The rate angle is gradually optimized to generate the network,so as to ensure the recovery of image clarity after processing.Through parameter selection and experimental research,after adding the self-attention layer and multi-scale discriminator,it is proved through experiments that the peak signal-to-noise ratio is increased by 0.17 and the structural similarity is increased by0.0374 in the rain image and the de-rain image.Comprehensive experimental results and comparative experiments,combined with self-attention and multi-scale generative confrontation network,have a better effect on raindrop removal.(2)Aiming at the problem of rain streaks attached to images,this paper proposes a method of feature fusion to generate a confrontation network.Since rain stripes are smaller in shape and more densely distributed than raindrops,it is easy to ignore the edges of rain stripes and detailed information in the process of removing rain stripes.First,the high-frequency information of the image is separated by wavelet transform,and the high-frequency features of the image are extracted and enhanced by the convolution and SE(Squeeze-and-Excitation)module,and then the high-frequency features and the features of the original rain stripe image extracted by the network are superimposed Fusion,so that the network can learn more accurately the rain streak feature and finally remove it and restore a clear image.Experiments show that the peak signal-to-noise ratio and structural similarity of the image are increased by 0.74 and 0.0233,respectively.Therefore,feature fusion is beneficial to the effective removal of rain streaks.For raindrop removal,this article uses a combination of self-attention generation network and multi-scale discriminant network.For rain streak removal,this article uses feature fusion to generate a confrontation network.This article summarizes and analyzes both quantitative and qualitative aspects to illustrate the improved generation confrontation proposed in this article.The Internet has improved the removal effect of raindrops and rain streaks on photos taken on rainy days,and the clarity of the image has been restored to a higher level.Since image rain removal is a preparatory step in the research image processing process,finally,this article builds an image preprocessing platform,including the image raindrop removal and image removal rain streak studied in this article,to realize the visualization operation of different models,and can display and display Save the rain results.