Research On Image Rain Removal Method Based On Dual Scale Analysis

Images are the main channel for people to obtain information.However,images taken in rainy weather will be seriously disturbed,which will hinder subsequent image analysis.Because rain streaks are very easy to cause a series of physical reactions such as light scattering and reflection,the digital image produced by the imaging system will greatly reduce the quality.For computer vision algorithms and related applications that work in outdoor scenes,such as target recognition,autonomous driving,and behavior detection,the objects they process are assumed to be based on clean images.On rainy days,their operation will not adapt to the situation,resulting in failure to meet market demand.Therefore,it is of great practical significance to restore rainy images into clean images with clear scenes.The rain-removal from images needs to take into account both the removal of rain streaks and the restoration of clear scenes.The results obtained by the existing method of removing rain from images are prone to the phenomenon that rain streaks can be removed but the scene is very blurred.In this paper,a novel two-scale single-image rain removal method based on image dual-scale analysis is proposed in the framework of deep learning based on this phenomenon.In summary,the core work of this article mainly includes the following three points:(1)A dual-scale analysis was performed on the causes of rain marks that would reduce the image quality.A guided filter can be used to separate the base layer from the image and obtain the detail layer.Through this method,this paper extracts the respective base layer and detail layer from a pair of synthetic rainy images and clean images in turn,and then performs histogram statistics on the RGB pixel values of these layers in the experiment.Analyze the essential reason of the impact caused by rain streaks.The above-mentioned dual-scale analysis can profoundly explain the essence of the impact of rain streaks on the image,and provide theoretical support for the dual-scale single-image rain-removal method proposed in this paper.(2)By introducing a squeeze-excitation neural network unit,a squeeze-excitation residual network for image rain-removal is proposed.Attention mechanism has been emerging in the field of natural language processing,so this paper introduces this mechanism into image deraining by using squeeze-stimulated neural network units.The squeeze-excitation neural network unit can have a filtering effect on the feature channels in the network learning process,so that the entire network model learns the rain streaks features more accurately,and finally improves the results of image rain-removal.(3)Combining the two-scale analysis of rainy images and the squeeze-excitation residual network,a dual-scale squeeze-excitation residual network is proposed,which can greatly improve the results of image rain-removal.By analyzing the effects of rain streaks on the image,this article explains that rain streaks damage the image's base layer and detail layer to varying degrees.To this end,the process of the image rain-removal is refined,and the convolutional neural network learns the base and detail layers of the corresponding clean image at the same time,and finally merges the two layers to learn again,and outputs a complete clean image.