Research On Image Rain Removal Method For 3d Modeling

The purpose of single image rain removal is to eliminate the rain line in the image and restore a clear background image.Compared with video image rain removal,single image lacks the inter frame information of the image,so it is more challenging.Rain streaks and rain lines brought by rainy days will reduce the quality of UAV aerial images,which will have a negative impact on a series of practical applications,including 3D modeling.Building high-quality 3D models requires high-quality image information input.Therefore,the study of single image rain removal and its application in 3D modeling are of great significance and value.On the one hand,in order to remove the adverse effects of rainy days on images and improve the quality and application value of images,we explored the unified framework of multi-scale representation from the input image scale and depth neural network representation,and proposed a multi-scale progressive fusion network(RACS Net)based on attention mechanism to interpret a single image using the relevant information of rain stripes of different scales;On the other hand,due to the great uncertainty of the 3D modeling effect of UAV images in rainy days,we first use the new rain removal network proposed in this paper to remove rain from rainy images,and then use 3D modeling software 3DS Max to model the three digit building,to verify the effectiveness of the rain removal network and the necessity of its application in the direction of 3D modeling.Considering the safety problem of UAV aerial photography in rainy days,the simulation data set is used in the 3D modeling stage of this paper.The main research work of this paper is as follows:(1)This paper proposes a rain removal network based on the mixed attention mechanism and residual gating module.Because the input image contains rich feature information,not every region and channel are equally important,so the mixed attention module is added to better extract the effective spatial information of the image,screen important feature channels,and further extract rich structural information through the residual network,Keep more details and restore image quality more clearly.(2)This paper proposes a feature extraction network based on multi-scale feature fusion mechanism,mixed attention mechanism and residual gating mechanism.Due to the large number,different sizes and different directions of rain lines in the image,it is difficult to effectively restore the image quality while removing rain lines.The module designs the convolutional layer network with different receptive field sizes,samples the images to different resolutions,removes rainwater information of different scales,and effectively fuses the global context information to further improve the rainwater removal performance.(3)In this paper,a large number of comparative experiments and ablation studies have been carried out on multiple benchmark datasets for the proposed single image rain removal method,which proves that the proposed method is more generalized in rain removal tasks and has high performance in rain removal and quality recovery.It is applied to the image rain removal of the simulated UAV aerial data set,and the data set after rain removal is applied to 3D modeling,which further verifies the effectiveness of the rain removal method on the authenticity of 3D modeling of images in rainy days.