Single Image Rain Removal Based On Improved Spatial Attentive CGAN

With the rapid development of computer vision technology,outdoor vision system has been widely used in military,traffic and safety monitoring fields.Many outdoor vision systems require accurate outdoor scene detection,such as outdoor video monitoring,driverless system,etc.However,in rainy days,the visibility of the scenery is greatly reduced,and the pictures and video data are greatly affected by raindrops.The image will become blurred so that the key information in the image can not be recognized,which will lead to the use value of the image greatly reduced.At present,most of the researches on rain removal are focused on video or multi frame images,mainly because of the lack of prior information such as rain drop sensitivity change and spatiotemporal correlation in a single image,which makes the problem of rain removal in a single image more complex.Based on the limitation of lack of prior information in a single image,there are two major difficulties in this technology: first,although rain lines can show thin and bright lines in the image,it is difficult to accurately detect complex and changeable rain lines due to wind direction factors,similar background interference and other problems;second,excessive or insufficient rain removal often occurs when removing rain lines,which will lead to the image fault,speckles,or the loss of image details after rain removal.In recent years,with the development of deep learning network,with its advantages in feature extraction,many scholars have made great progress in the task of removing rain from a single image.However,there are still some problems in the deep learning method used in the task of removing rain from a single image.First,because the real rain and no rain image data pair can not be obtained,and the synthetic data set is difficult to cover the complex and diverse rainfall distribution in the real rain image.At present,the single image method can only remove one kind of rain line contained in the training data set effectively,and the generalization ability of the model is low,which is not conducive to practical application.Therefore,it is necessary to improve the generalization ability of the model for rain line recognition to improve the rain removal effect of a single image,so as to improve the practical application value of the model.Secondly,the results of single image after rain removal model often appear the phenomenon of variegated spots,image contrast saturation change and even unsmooth rain line after rain removal.Therefore,aiming at theabove two problems,this paper proposes a high-quality single image rain removal model with high generality,taking CGAN(Conditional Generative Adversarial Networks)as the baseline network,and then makes the following two improvements,improving the generalization ability of the model and enhancing the quality of the image after rain removal:(1)Aiming at the problem that the current single image rain removal method has insufficient generalization ability for rain line recognition,this paper first adds the improved spatial attention module on the basis of CGAN,constructs a general model which can be applied to multiple rain line removal,and improves the practical application value of the model.After the first improvement,through the improved spatial attention mechanism to enhance the ability of recognition and location of rain line,make the model focus on the location of rain line,reduce the interference of other background information,and solve the problem that only one rain line can be effectively removed by the current single image method.Experimental results show that the single image rain removal model with improved spatial attention can effectively remove rain in multiple rain line situations on the premise of only using one rain line data set for network training,which verifies the effectiveness of the single image rain removal model with improved spatial attention to improve the universality of the model.And compared with other mainstream algorithms,the method in this paper gets better results on the most efficient and realistic public dataset.(2)Aiming at the problem that the detail restoration effect of the existing single image rain removal model is not good after rain removal,based on the addition of the improved spatial attention module,the second improvement of the model is carried out,and the multi-scale weighted fusion module is added,so that the method proposed in this paper can further enhance the quality of the image after rain removal on the basis of improving the universality of the model.After further improvement of the model,through multi-scale weighted fusion module to improve the ability of model feature extraction,assisted by the improved spatial attention mechanism module to further enhance the recognition and positioning of rain lines,while removing rain lines,better retain image details,improve the quality of image generation.At present,it improves the rain removal effect of the model on the most realistic public dataset,and improves the recovery ability of image details while effectively removing rain on other public datasets,which verifies the effectiveness of themodel in improving the generalization ability and improving the image quality after rain removal.At the same time,the effect of rain removal is better than other methods on the real rain images,which proves that the method proposed in this paper has high practical application value.