Related Experiment Video
Updated: May 14, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
Discrepancy-Guided Semantic Segmentation with Boundary Detail Enhancement for Traffic Scenes
Changshun Yu1, Xiujian Yang1,2, Shiquan Shen2
1Faculty of Mechanical and Electrical Engineering, Kunming University of Science and Technology, Kunming 650500, China.
Sensors (Basel, Switzerland)
|May 13, 2026
Summary
This study introduces a novel method for traffic scene semantic segmentation, enhancing fine-grained object detection and boundary accuracy. The approach improves detail preservation and fusion robustness, leading to better overall segmentation performance.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Traffic scene semantic segmentation faces challenges with fine-grained objects, blurred boundaries, and detail suppression during cross-scale fusion.
- Existing methods struggle to effectively integrate multi-scale features while preserving crucial shallow details and accurate boundary information.
Purpose of the Study:
- To propose a discrepancy-guided semantic segmentation method with boundary detail enhancement for traffic scenes.
- To improve the semantic completeness of fine-grained regions and alleviate boundary ambiguity and detail loss.
Main Methods:
- Introduced a Gated Collaborative Context Module (GCCM) for adaptive multi-scale semantic dependency capture.
- Designed a Frequency-Edge Guided Enhancement Module (FEGE) to enhance boundary details using frequency decomposition and edge operators.
- Proposed a Discrepancy-aware Pixel-Adaptive Gating Fusion module (D-PagFM) for robust cross-scale feature fusion.
Main Results:
- Achieved mIoU scores of 80.08% on Cityscapes and 82.97% on CamVid datasets.
- Demonstrated significant improvements in boundary-sensitive categories like road boundaries, poles, and traffic signs.
- Showcased enhanced robustness in feature fusion and boundary consistency.
Conclusions:
- The proposed method effectively addresses challenges in traffic scene semantic segmentation, particularly for fine-grained details and boundaries.
- The integration of GCCM, FEGE, and D-PagFM modules leads to superior performance and high-precision segmentation.
- The method holds significant potential for real-world applications requiring accurate traffic scene understanding.