Rain Removal With Single Image Based On Deep Convolutional Neural Networks

Under rainy weather,especially in extreme rainy weather such as heavy rain and torrential rain,the images captured by outdoor computer vision systems can produce severe fog blurring,which affects their application performance.Using image processing techniques to eliminate the effect of rain and fog blur and improve the clarity of the computer vision system image picture is of great significance and practical application value to enhance the application capability of computer vision system to adapt to rainy weather,which is a research hotspot in the field of image processing and computer vision in recent years.In this thesis,Research on the deep learning method of removing rain from images.The main work and achievements are as follows:(1)Aiming at the problems of obvious residual rain marks and loss of local details in the existing image deraining deep convolutional neural network methods,a rain images removal method based on the residual dense expanded convolutional connection network is proposed.The overall framework consists of a rain pattern removal network and an image enhancement network.Among them,the rain pattern removal subnet structure adopts the encoding and decoding structure,and uses the residual dense connection module as the core building unit.The image enhancement network is composed of multiple feature extraction modules with the same structure based of different expansion rate convolution operators to ensure that the feature information about different receptive fields can be fully extracted and more image details can be restored.Experimental results show that the proposed method can effectively remove rain marks and restore image details.Compared with the existing methods,the PSNR and SSIM is improved by 4.96dB and 0.20 on the Rain 100H dataset,and a better visual effect is obtained.(2)Aiming at the long running time of existing image deraining deep convolutional network methods,a neural network method of rain removal of image based on attention and the ensemble of context information is proposed.This method uses the encoding and decoding network as the basic framework.In order to be able to extract features of different levels,the coding network extracts image feature information by aggregating context features.In order to increase the correlation between channel features and improve the processing speed of feature aggregation,an attention mechanism is introduced in the process of convolution operation.The experimental results show that the method of this paper can effectively remove the rain pattern and shorten the running time.Compare with the comparison method,in the size of 512×512 images,the average processing time of a single image is shortened by up to 0.2 seconds.(3)Based on the above theoretical research results,using PyCharm and Pyqt5 to design an image rain removal prototype system based on deep convolutional network,the system can selectively remove rain from the image according to needs.