Semantic segmentation network of remote sensing image buildings based on LEDNet

As an important product and component of social development, buildings are closely related to various activities of human society. In this study, LEDNet network is used to extract buildings in order to solve the problems of imprecision edge extraction and misclassification. The encoder-decoder structure is adopted in this method. The encoder uses ResNet as the backbone network, channel split and shuffle operations are introduced into each residual block, and the attention pyramid network APN is adopted in the decoder. While reducing the complexity of the whole network and greatly reducing the computing cost, it can also maintain high segmentation accuracy. The evaluation indexes F1 and IoU on the WHU building dataset reached 93.54% and 87.86% respectively, and all evaluation indexes were higher than the other four comparison algorithms, which verified the advantages of the new method.