Research On Single Image De-raining Based On Convolutional Neural Network

Image de-raining refers to the removal of rain in the image to restore a high-quality image,which belongs to the underlying visual image processing.In recent years,with the rise and wide application of deep learning and imaging,image de-raining has also attracted the attention of many scholars.This article mainly takes convolutional neural network as the basic theory.Combined with the latest expansion convolution,attention mechanism,dual path network and gated network,etc.,a targeted network structure model is designed to achieve an effective single image rain removal task.The specific research work of this article is as follows:(1)Single image de-raining using a recurrent dual-attention-residual ensemble network is proposed(RDARENet).Aiming at the problems of the current single image de-raining algorithm,such as rain streaks remain and image detail texture restoration effect is not good,this paper proposes a new single image de-raining using a recurrent dual-attention-residual ensemble network.By using the original rain image as a global feature,the network is concatenated with the output image of the rain removal module at each stage to gradually remove the rain streaks,and a dual attention mechanism is added to better distinguish the rain streaks and the background layer.Finally,a gated network is used in the output module of the network to automatically fuse image detail information at various stages to obtain high-quality images.(2)A multi-scale encoding-decoding model for single image de-raining is proposed(MSEDNet).Based on the consideration of computing time for single image de-raining,this paper proposes a single image de-raining network with multi-scale encoding-decoding structure.Drawing on the idea of the encoding-decoding structure,the rain image is down-sampled in the encoded structure.Each down-sampling layer is separately operated by the feature refinement module to extract rich feature information.Finally,the feature map is up-sampled step by step,and aggregated with the feature map of the previous stage.The feature extraction network at different scales can better aggregate the features of rain streaks.Experimental results on simulation and real rain data sets show that the proposed network is very good at removing rain streaks and has a faster calculation speed.