Research And Implementation Of Image Deraining Based On Deep Learning

Rainy weather is a common weather condition,which not only seriously affects human visual perception,but also degrades the performance of various high-level vision tasks,such as image classification,and video surveillance.This article starts from the needs of actual application scenarios,aims at the pain points of engineering applications,and recovers clear scene images from rainy images with different levels of interference.In order to effectively remove the image rain and fog blur interference in the computer vision task system and improve the clarity of the image,this paper analyzes and studies the rain mark images under different environmental conditions,and develops an efficient and reliable image deraining network based on generative confrontation,aiming at It provides clear images for the relevant links of advanced visual tasks of intelligent systems such as smart construction sites and video surveillance.This paper focuses on the creation of a more realistic rainy image dataset and the improvement of the performance and efficiency of the image removal algorithm.The main research content is as follows:In order to solve the problem of ignoring fog occlusion and scene depth information in the existing synthetic rain image data set,this paper first established a rain mark image model based on the principle of atmospheric scattering,and analyzed the scene depth in the actual vision to calculate the depth estimation map.Synthesize a rainy image data set that is more suitable for real scenes,and tap space for improving the generalization ability of the rain mark removal model.Secondly,in order to further optimize the deraining performance of the generator,an image deraining network based on generative adversarial technology is proposed.Among them,the frequency decomposition module is convenient to narrow the mapping range of network learning,and reduce the burden for the subsequent work of removing rain marks;the generator module repeatedly expands the shallow Res Net,and uses the output of the previous stage and the original rainy image as the input of the next stage.The network parameters are significantly reduced without reducing the performance of rain marks removal,and the recurrent layer and attention mechanism are further introduced,so that the generator can fully retain the background texture information of the image while gradually removing the depth features of rain marks in the image;The discriminator module discriminates the authenticity of the derained images generated by the generator.Both a single MSE and a negative SSIM loss are applied to train the network.The research results show that the deep learning model of this research performs well in both synthetic and real rainy images.Compared with other algorithms,the PSNR and SSIM of the model are increased by 4.96 and 0.20 respectively,and the processing time is excellent,and the average time is shortened by 0.2 Second.Based on the above theoretical research results,this paper designs and implements an image removal and sharpening system based on the Pyqt5 framework.This system can effectively remove rain marks and improve image clarity,and has certain practical application value.