Research On Single Image Rain Removal Based On Deep Learning

The rainy image can be expressed as a linear superposition of the clean background part and the rain streaks.Single image rain removal is an ill posed layer segmentation problem,which aims to segment the clean background from the rainy image.The traditional algorithm of removing rain from a single image has high cost,and there are some problems such as color distortion and detail loss in the recovered image.In recent years,the method based on deep learning has been widely used in the field of image.In order to improve the performance of the image rain removal algorithm,In this paper,convolution neural network is used to further study and analyze the problem.The specific research contents are as follows:Firstly,a fast single image rain removal algorithm based on dual recursive fractal network is proposed.On the one hand,in order not to increase the network parameters,fractal network recursion many times,sharing parameters,reducing the training difficulty of the network.The fractal network chooses smoothed dilated convolution with different dilated factors to increase the receptive field of the network,so that the feature map contains more scene information.On the other hand,the designed network removes rain from the output rainy image stage by stage,so as to improve the performance of the network.The experimental results show that the algorithm can effectively remove the rain streaks from the synthetic and real rainy images.Secondly,in order to reduce the impact of the lightweight network in the rapid rain removal at the cost of reducing the effect of image rain removal.A single image rain removal algorithm based on hierarchical multi-scale residual network is proposed.In this algorithm,convolution kernels of different sizes are used to improve the ability of network learning image details,and more complex features are learned by appropriately increasing network depth through hierarchical structure.Residual learning effectively relieves the problem of gradient disappearance brought by deep network.The experimental results show that the restored image can retain the background details of the image while removing the rain,and the overall visual effect is more clean and nature.Finally,in order to retain the color information of the source image and improve the color saturation of the restored image,a single image rain removal algorithm based on the hierarchical cross network of channel decomposition and fusion is proposed.The algorithm decomposes the YUV color space of the image.Deep network is used to learn the brightness channel which is greatly affected by rain,while shallow network is used to learn the color component which is less affected by rain.Finally,the reconstructed YUV spatial image is input into the fusion sub-network to get the restored image.The experimental results show that the algorithm can effectively remove the rain streaks and retain the details and color information of the source image.