Research On Single Image Rain Removal Based On Convolutional Neural Network

With the progress of digital technology and the vigorous development of image processing technology,outdoor computer vision systems have been widely used in real life,such as aerial photography,monitor and automatic navigation.However,videos and images of outdoor scenes are often interfered by bad weather,which greatly limits the normal use of computer vision systems.Besides,the performance of computer vision algorithms such as object detection and location will also be affected by image quality.Rainy day is one of the most common weather in nature.It is difficult to get a clear image when shooting the outdoor scenes on a rainy day.The image quality decreases and becomes blurred when rain streaks cover the objects in images,which has a unwanted impact on the accuracy of computer algorithm.Therefore,it is of great significance and practical value to carry out the research of single image rain removal.In this paper,the convolutional neural network is used as the research tool to study the task of single image rain removal.By reasonably designing the networks,good performance of rain removal and fast computation speed are both achieved.The specific research results and contents are summarized as follows:1.A simplified residual dense de-raining network is proposed(SRDN)In order to solve the problems of poor de-raining performance and long operation time in the advanced single image de-raining algorithms,in this paper,a simplified residual dense de-raining network is proposed.The proposed network includes an improved residual connection and a simplified dense connection block.By using image decomposition and residual knowledge appropriately,the residual connection can make the network training easier and accurately learn the luminance information between rainy images and clear images.The simple dense network has two advantages: firstly,by dense connection,it can effectively use the features to train the network,which is beneficial to retain more details in derained images;secondly,it can greatly improve the computation speed by simplifying the network module and using fewer network parameters.The experimental results show that the proposed simplified residual dense network is superior to many advanced single image de-raining algorithms,and has good de-rainingperformance and high computation speed.2.A weighted residual de-raining network is proposed(WRN)To further reduce the computation time while maintaining good de-raining performance,in this paper,a weighted residual rain removal network is proposed.First,a weighted residual connection mode is designed by introducing the weighted idea into the above-mentioned residual connection of SRDN,so that the network can dynamically adjust the use of residual features in the learning process.Then,base on the convolution-deconvolution network with skip connection,a multi-scale weighted network module is proposed.The module has three advantages: Firstly,the multi-scale convolution is used to increase the receptive field of the network and promote the information fusion;Secondly,the weighted idea is applied to parameter layer to dynamically adjust the contribution of network layers to the whole network;Thirdly,the convolution and deconvolution both with stride 2 are used to largely reduce the computation time.The experimental results prove that the proposed network achieves high de-raining performance,especially in the computation speed,it exceeds several advanced single image de-raining algorithms.