Research on Remote Sensing Image Segmentation Network Based on Improved LRASPPHead

The study aiming at the problems of rich targets, uneven distribution of targets, large degree of occlusion between targets, and high similarity between different buildings in remote sensing images related to buildings in urban blocks, the deep recognition technology is used to study the semantic segmentation of buildings, streets and vegetation in cities. The article is based on the LRASPP deep learning semantic segmentation network. By improving the network structure of each layer, the mean Intersection-over-Union and Pixel accuracy of semantic segmentation are improved on the premise of that a certain training speed is maintained. And more accurate segmentation results are obtained. The network in this paper sheds some light on the semantic segmentation of images with rich and complex details, images with small differences between target and non-target objects, and images with variable target forms.