Related Experiment Video
Updated: Apr 2, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.7K
BGSC-Net: Boundary-guided semantic compensation network for remote sensing image segmentation.
Xin Wang1,2, Zhe Lu3, Qun Yang2
1School of Computer and Information Security, Guilin University of Electronic Technology, Guilin, China.
Plos One
|March 31, 2026
Summary
The Boundary-Guided Semantic Compensation Network (BGSC-Net) improves remote sensing image segmentation by enhancing feature fusion and boundary guidance, leading to better detection of small objects and detailed structures.
Area of Science:
- Computer Vision
- Remote Sensing
- Deep Learning
Background:
- Deep learning, particularly hybrid CNN-Transformer models, shows promise for remote sensing image segmentation.
- Challenges persist in complex scenes, including capturing fine boundary structures and small objects due to suboptimal feature fusion and lack of explicit boundary guidance.
Purpose of the Study:
- To propose a novel hybrid architecture, BGSC-Net, to address limitations in remote sensing image segmentation.
- To enhance the segmentation of small objects and fine structural details in complex remote sensing scenes.
Main Methods:
- Introduced the Boundary-Guided Semantic Compensation Network (BGSC-Net).
- Integrated a Cross-Level Semantic Compensation Module (CLSCM) for dynamic fusion of semantic and spatial details.
- Incorporated an Auxiliary Boundary Supervision Module (ABSM) for explicit boundary modeling and joint optimization.
Main Results:
- BGSC-Net achieved superior segmentation performance across multiple datasets: Potsdam (87.57% mIoU), Vaihingen (85.61% mIoU), LoveDA (55.05% mIoU), and UAVid (74.77% mIoU).
- Demonstrated strong generalization on a specialized fine-grained task, achieving 89.58% mIoU on the Mangrove Species Fine-grained Segmentation Dataset (MSFSD).
Conclusions:
- BGSC-Net effectively overcomes semantic misalignment and edge information loss in remote sensing image segmentation.
- The proposed model shows practical utility for precise mapping tasks, including fine-grained species identification.
