Application Research Of TV Model In Raindrop Removal From A Single Image

Rain is a common bad weather phenomenon.Because of the interference of raindrops,the quality of images collected in rainy days will deteriorate,which seriously affects the subsequent work like image segmentation and target recognition.Therefore,the methods of raindrops removal in images has became an important research content in the field of computer vision and artificial intelligence.In this thesis,the raindrop removal method of single image is studied.To overcome the shortcomings of current methods of raindrop removal such as filtering and so on,which can only remove the Rain line stripes in images,the deep learning method is too dependent on samples and the training process is complex.On the basis of the physical modeling of raindrops,the classical TV model of image restoration is introduced to process raindrop images.Combined with the improved region restoration method,the image after raindrop removal can be clarified.In this thesis,the structure-texture decomposition TV model such as TV-L2,TV-L1,TV-G,TV-Hilbert,TV-Gabor,RTV are introduced,and the application scope,advantages and disadvantages of each model are analyzed.The differences of extraction ability of different types of models for texture and structure parts of images are emphatically analyzed.In this thesis,the characteristics of raindrops are analyzed.The physical model of raindrops is established by using spherical harmonic disturbance.The raindrops images are classified according to the relationship between the diameter and shape of raindrops,The physical models of raindrops are combined with TV-Gabor model and RTV model respectively,and new TV-Gabor-R and RTV-R models are obtained.The new models solve the defect that the original models cannot estimate the texture parameters of the model,but can only be determined by many experiments.The two models are used to design and implement the raindrop removal algorithm for a single image.The experimental results show that TV-Gabor-R model has a strong ability to remove raindrops in images,but it will cause a slight decline in image quality;RTV-R model has a weak ability to remove raindrops in images,but the image quality after raindrops removal is higher.In this thesis,in order to enhance the ability of removing raindrops from a single image,uses region restoration to process the image after removing raindrops.Aiming at the problem that region restoration is prone to"scar",the region restoration method is improved,and the improved method is combined with TV-Gabor-R and RTV-R models respectively to remove raindrops from images.The experimental results show that the proposed method is not only suitable for images with different raindrop sizes,but also can eliminate the residual raindrops and a small amount of reflected light in the images,and improve the image clarity.And the experimental data set is constructed by network collection and actual shooting,and the validity of this method is verified by peak signal-to-noise ratio.