Research On Image Rain Removal Algorithm Based On Deep Learning

Outdoor computer vision tasks are affected by bad weather such as fog,rain,snow,dust and so on.The image quality of the images collected is usually not high,which seriously affects the execution of computer vision tasks.The most common bad weather is rainy day,and the rain removal of single image is different from the rain removal of video.In a single image,it is usually difficult to distinguish the rain line from other textures,resulting in the removal of part of the background information,and the rain removal of single image cannot obtain useful information according to the change law of time dimension.This paper focuses on the research of single image rain removal.Currently,there are two main categories of single image rain removal methods,which are generation adduction network based rain removal and deep neural network based rain removal.There are two problems in generating adversarial network to derain.One is that the training time of generating adversarial network is long and requires a large amount of computing resources.The other is that the training is unstable and the generalization ability is not good,so it is easy to generate images with too high similarity.The common problem of the rain removal method based on deep neural network design is that although the effect of rain removal is beneficial when the number of network layers is deepened,the time of rain removal is extended correspondingly with the deepening of the number of network layers,and the timeliness is low.At the same time,when the direction or texture of the rain line is similar to the image background,part of the background and rain line are removed at the same time.Based on the problems of the above methods,two methods of rain removal based on generative adduction network and deep neural network are respectively proposed in this paper.The main work is as follows:(1)Then this paper proposes an image deraining algorithm based on multi-scale attention generative adversarial network according to the goal of image deraining task.This model is based on the GAN confrontation idea,and continuously improves the image generation quality of the generator through the feedback of the discriminator until it reaches a balance.In the generation network,the image features are extracted based on the residual network,and the SE attention mechanism is introduced,and multi-scale feature fusion is used to obtain more image background information.The encoder-decoder structure uses Res U-Net,which can better retain image features;In the discriminative network,a multi-scale discriminator is used to consider information of different scales at the same time,which can better capture the detailed features of the input image,thereby improving the performance of the model.Finally,full comparison experiment and ablation experiment were carried out on the public data set.The experimental results show that the proposed method has a good effect on removing the rainline in the image.(2)The above algorithm has a good effect in removing small rain lines,but the generative adversarial network has problems such as slow training speed and difficulty in convergence,and when removing dense and complex rain lines,it will still leave traces of rain removal.According to the above two problems,this paper further proposes an image deraining algorithm based on residual dense network.The algorithm includes a feature extraction module based on the Rep VGG network,which can ensure computational efficiency while saving memory;at the same time,an image enhancement network based on Dense Net is designed,which uses dilated convolution and CBAM attention mechanism internally to obtain multi-scale information and enhance the ability of network learning features,and in the reasoning stage,the multiway structure can be changed into a single structure,which increases the reasoning speed and reduces the memory consumption.A comparison experiment was conducted between the rain removal method and a number of deep learning-based rain removal methods on a public data set.The results show that the rain removal method has a good effect in removing image rain lines,and the rationality of the method was proved by the ablation experiment.