Single Image De-rain Based On Deep Neural Network

In many applications,such as video surveillance,driverless cars,outdoor sports events,etc.,the problem of image and video unclear caused by rainy conditions poses great challenges to image restoration,which greatly limits the limitations.The performance of computer vision algorithms.Therefore,it is important to develop algorithms that automatically remove these effects on image quality.This thesis discusses the problem of removing rain streaks from a single image.Various methods have been proposed before,but there are some drawbacks.One of the main limitations of the existing single image removal methods is that they are designed to process a certain types rainy images of some rainy conditions,and they do not effectively take into account various shapes,scales and raindrop densities in their algorithms.In order to solve these problems,two different single-mode rain removal methods based on the deep learning theory are studied and proposed which achieve good performances.The main idea of this thesis is as follows:Firstly,this thesis proposes a new dataset based on the conditional generation adversarial network to be more realistic.It combines the traditional image processing method to simulate the rain density,and the rain image dataset contains different rain direction and size.Through these methods of artificially synthesizing data sets,a rich sample is provided for rain training.Although the rain scene pictures of the training network are synthetic,they can be well trained to the proposed network,and can be well applied to real rainy images.Secondly,this thesis proposes a single image rain removal method based on HSI image space,which maps rain source images from RGB space to HSI space,and then learns the rain image and no rain image features based on DenseNet.The deep residual feature map of the rain image is finally superimposed on the luminance component and the residual feature map of the source image to obtain the reconstructed rain image,and inversely maps it to the RGB space,which preserves the chrominance information of the image well.Finally,an image-based rain-removal method based on density perception is proposed.This method uses a multi-stream DenseNet to remove rain and automatically determines the rain density information of the input rain image.In order to accurately estimate the density level of rain,this thesis proposes a new residual recognition classifier that uses the residual components in the rainy day image for density classification.Once the rain density level is estimated,we fuse the estimated density information and use the multi-stream DenseNet to get the ultimate rain output.In addition,ablation studies are performed to demonstrate the role of the different modules in the proposed network.