Research On Single Image Rain Removal Based On Deep Learning

Single image rain removal aims to remove the rain streaks from the rainy image and restore a clean image.However,the existing single image rain removal algorithms tend to remain rain streaks and remove some background details.In recent years,deep learning has achieved remarkable results in many image processing and computer vision tasks.In order to improve the performance of the image rain removal algorithm,this paper applies the deep learning method to the single image rain removal problem.The main research contents are as follows:Firstly,a single image rain removal algorithm based on PolyNet is proposed.This algorithm designs PolyInception modules and then uses them to build a single image rain removal network structure.It achieves the nonlinear mapping between rainy and clean images in an end-to-end manner.This algorithm uses the various forms of polynomial combination of Inception residual units to build the different PolyInception module.This design can increase the performance of network by enhancing the structural diversity.The experimental results show that the proposed algorithm can remove rain streaks in different directions and densities from rainy images and preserve image background information.Secondly,in order to further improve the performance of the single image rain removal algorithm,this paper uses the idea of densely connection to construct a single image rain removal network structure based on DenseNet.This dense connectivity pattern can strengthen feature propagation and alleviate the gradient vanishing problem during training.The network uses dilated convolution to increase the receptive field.The experimental results show that the method improves the image de-raining effect and makes the restored image have better visual quality.Finally,in order to make the restored image retain more accurate scene detail information,this paper proposes a single image rain removal algorithm based on image decomposition and a dense network by using the information that rain component belong to the high frequency.The two sub-networks are designed to learn the low-frequency part and the non-rain high-frequency part of the rainy image respectively,and then the learned low-frequency image and the high-frequency image are superimposed to obtain a restored image.This algorithm add total variation regularization term and LF-channel fidelity term to the loss function to optimize the two sub-networks jointly.Experiments show that the algorithm can remove rain streaks in the image while retaining more accurate scene details.