Research On Single Image Rain Removal Based On Non-local And Fuzzy Width Learning

With the gradual development of computer vision technology,various kinds of outdoor computer vision related tasks are gradually coming to people's visions.Precipitation weather is one of the important factors affecting the effectiveness of outdoor visual system.Therefore,image rain removal technology has gradually become a research hotspot.These technologies can be widely used in outdoor monitoring,unmanned vehicles,remote sensing and other fields.Single image rain removal is a branch of rain removal research.Compared with video rain removal,which can make use of redundant multi-frame information,single image rain removal is more challenging.This paper discusses the problem of single image rain removal and puts forward a new solution.The core idea of this paper is to decompose the rain water removal task in a single rainy image into the distant rain removal task and the close-up rain removal task in the rainy image,and apply the non-local dehazing technology and the fuzzy broad learning technology to the near and far rain removal task respectively.The main research contents of this paper are as follows:1.In this paper,firstly,the dehazing algorithm is used to remove the perspective rain from the input rainy image.In view of the phenomenon that the background reduction degree is not high and the background color is too dark after the traditional method of removing the perspective rainwater,this paper applies the non-local dehazing technology to the subject of removing the rain in a single image,and removes the distant rainwater in a single image.This technology can remove the distant haze and rain in the rain weather and highlight the rain streaks in the near view.2.Then,in the process of dealing with the rain in close range,this paper designs a rain removal network based on fuzzy broad learning system,aiming at the problems of too many hyper-parameters of deep learning method and too long training time.In the pre-processing of training data,the image is divided into detail layer and base layer,and the training data is converted from the detail layer of RGB space to the Y channel image of YCbCr space,which greatly reduces the amount of training data and speeds up the network training speed.3.Finally,the details extracted from the dehazed image are superimposed on the preliminary results with a certain degree of transparency to obtain the final results.Experimental results based on real world images and composited images show that the proposed solution has good performance while reducing running time and training time.