Study On Image Deraining Algorithm Based On Recursive Neural Networks

Rainy weather will introduce rain stripes into outdoor images or videos captured by surveillance cameras.Rain strips can block the image background and make the image blurred.Generally speaking,image definition is directly related to the reliability of image-based application systems.With the increasing demand of computer vision tasks,such as target detection,face recognition and automatic driving,the rain removal processing of rainy day images has a broad market application prospect.With the continuous development of deep learning theory and technology,methods based on deep learning have natural advantages in automatic extraction of image features using deep networks,and can learn complex nonlinear mappings between rain images and clean backgrounds.The rain image is usually formed by accumulating rainy stripes of varying degrees and varying directions.Most of the existing methods remove the rain strips in multiple stages,without fully considering and utilizing the information transfer between stages.In the case of complex rain scenes,insufficient rain removal is easy to occur.Therefore,this paper focuses on the image deraining algorithm based on recursive neural network.The main work is as follows:(1)An image deraining algorithm based on multi-scale hierarchical structure of recursive neural network is proposed.Since the rainy image contains rain strips with different sizes and directions,it is sampled to different resolutions,and auto-encoder networks,with down-sampling operations of different sampling rates,are used to remove rain stripes with corresponding different intensities.Since the context information is very crucial for rain removal,the long-short term memory network is used to retain the useful information in the previous stage and guide the rain removal in the later stage in the rain removal tasks with different resolutions.In addition,the attention mechanism SENet module is introduced to highlight important feature information and further enhance the rain removal effect of the model.The effectiveness of the proposed method is verified by comparing with state-of-the-art deraining algorithms.(2)A recursive neural network image deraining algorithm based on recurrent mechanism is proposed.In order to make the network more lightweight,based on the multi-scale hierarchical structure of the previous recursive neural network,the structure of rain removal network is improved.A full scale jump connection is introduced in the auto-encoder network,which changes the network connectivity and allows a fast training.Specifically,the recursive neural network image deraining algorithm based on recurrent mechanism is implemented by repeatedly expanding the auto-encoder and the long-short term memory network.Using the cross-stage dependence of features,a recurrent layer is introduced to form a multilevel recursive network.A large number of comparative experiments and ablation studies have been conducted on public datasets.Experimental results show that under complex background,the proposed method performs well in image deraining.