High-Resolution Boundary-Constrained and Context-Enhanced Network for Remote Sensing Image Segmentation

Remote Sensing

Jie Jiang¹  Yizhe Xu¹ 

[1] School of Instrumentation and Optoelectronic Engineering, Beihang University, Beijing 100191, China;

关键词: remote sensing image;    semantic segmentation;    attention mechanism;    boundary information;   

DOI  :  10.3390/rs14081859

来源: DOAJ

【 摘 要 】

The technology of remote sensing image segmentation has made great progress in recent years. However, there are still several challenges which need to be addressed (e.g., ground objects blocked by shadows, higher intra-class variance and lower inter-class variance). In this paper, we propose a novel high-resolution boundary-constrained and context-enhanced network (HBCNet), which combines boundary information to supervise network training and utilizes the semantic information of categories with the regional feature presentations to improve final segmentation accuracy. On the one hand, we design the boundary-constrained module (BCM) and form the parallel boundary segmentation branch, which outputs the boundary segmentation results and supervises the network training simultaneously. On the other hand, we also devise a context-enhanced module (CEM), which integrates the self-attention mechanism to advance the semantic correlation between pixels of the same category. The two modules are independent and can be directly embedded in the main segmentation network to promote performance. Extensive experiments were conducted using the ISPRS Vahingen and Potsdam benchmarks. The mean F1 score (m-F1) of our model reached 91.32% and 93.38%, respectively, which exceeds most existing CNN-based models and represents state-of-the-art results.

【 授权许可】

Unknown