Remote sensing building image extraction method based on improved U-Net model

• Main Authors: CHEN BIN, SHAO YANLI, WEI DAN, LI XIN, WANG XINGQI, FANG JINGLONG
• Format: Patent
• Language: Chinese; English
• Published: 24.03.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention provides a remote sensing building image extraction method based on an improved U-Net model. A U-Net model is used as a backbone network, and an attention coordination gating module is added in a jump connection stage of the model so as to better position building subject information and effectively integrate different levels of semantic information. In a bridge connection part of the model, an efficient continuous pyramid module is introduced, feature extraction is performed by using grouped small-scale cavity convolution, and feature layered superposition is performed in groups, so that the ability of the model to capture possible spatial context relationships among building main bodies is enhanced. In a decoder part of the model, a bidirectional cascade depth supervision module is introduced, and prediction maps obtained in adjacent decoder stages are sequentially fused in two directions from deep to shallow and from shallow to deep, so that specific supervision on different network layers is