Research On Single Image Rain Removal Method Based On Deep Learning

As common natural weather,rainy days will affect the imaging quality of outdoor vision systems and restrict the performance of subsequent advanced computer vision tasks.Therefore,the in-depth study of image rain removal has practical application value.Rain streaks are irregularly distributed in the image,which increases the processing difficulty of traditional model-driven methods.Due to the high difficulty of collecting real rainy data sets,most deep learning-based methods use synthetic data sets to train models,but synthetic data sets can’t represent the feature distribution of real rainy images,which limits the generalization of deep learning models.In addition,the existing deep learning rain removal algorithms have the defects of a large number of parameters and high computational consumption,which are difficult to be applied to embedded mobile devices.Therefore,based on the above limitations,this paper improves the existing deep learning rain removal algorithm from the two aspects of improving rain removal efficiency and model lightweight.The main work is as follows:(1)Aiming at the problems of artifact residual,loss of details,and slow processing speed in the current rain removal algorithm,an efficient multi-scale convolutional selfattention single image rain removal method(EMT-Net)is designed.This method combines high-efficiency convolution and self-attention,establishes the interdependence of near and far features and makes up for the global feature modelling ability that traditional convolutional networks can’t satisfy.The multi-scale spatial feature fusion module is embedded to effectively increase the receptive field of the network and realize the feature extraction of rain streaks of different sizes.A hybrid loss function is designed,which uses the advantages of each loss function to make up for the defects of a single loss function and enhance the detail recovery ability of the network.The experimental results show that the EMT-Net effectively removes rain streaks with different densities and directions,fully retains background details and has good generalization on real data sets,and the processing speed is significantly improved.(2)Aiming at the problems of a large number of parameters and high computational consumption in existing deep learning-based rain removal algorithms,a lightweight single image rain removal algorithm(MoU-Net)is proposed that integrates U-Net and improved MobileNet-V3-Large.The algorithm draws on the idea of the encoder-decoder framework of U-Net,using the improved MobileNet-V3-Large as the encoder for feature extraction,and replaces the standard convolution with the depthwise separable convolution,effectively reducing the number of parameters in the encoding stage and the amount of computation.The attention mechanism based on normalization is used to replace the channel attention of the original SE,and a weight sparse penalty is applied to the attention,which improves the computational efficiency while maintaining similar performance.Multi-scale spatial feature fusion module is added to improve the performance of different feature learning capability for density rain streaks.Rescale the scale factor in the mixing loss function to better preserve background details.The experimental results show that MoU-Net can still achieve a good rain removal effect under the premise of fewer parameters and calculations.