The real-time perception requirements for surface deformation and moving targets in civil engineering disaster monitoring are increasing day by day. This paper proposes an image processing method for unmanned aerial vehicle (UAV) oblique photography that integrates multi-view 3D modeling and an improved deep learning network. A high-precision three-dimensional surface model is constructed through multi-view image fusion, point cloud filtering optimization and error calibration techniques. Combined with the improved DeeplabV+ network, an image segmentation model including multiple modules such as the encoding network, spatial pyramid module and decoding network is constructed to achieve the accurate segmentation of landslide targets. The results show that the accuracy of the method proposed in this paper in the image processing of objects related to different civil engineering disasters reaches 86.57%, 88.54%, 92.32%, 88.46%, and 89.75% respectively, which is much higher than that of the comparison methods. In disaster monitoring, the application of the method in this paper can increase the identification rate of hidden danger points to 98%, advance the early warning time by an average of 8 days, and reduce economic losses by 1.6 million yuan.
