Related Experiment Videos
Lightweight Semantic Perception from UAV-Borne Visual Sensors via Conflict-Suppressed Heterogeneous Expert
Feng Ouyang1,2, Yongpeng Ding1,2, Miao Qin1,2
1College of Intelligent Technology, Tianfu College of Southwestern University of Finance and Economics, Mianyang 621000, China.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces MEKD-UAVSeg, a lightweight framework for accurate aerial image analysis using Unmanned Aerial Vehicle (UAV) visual data. It enhances semantic perception by distilling knowledge from specialized AI experts, improving efficiency for real-world applications.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Remote Sensing
Background:
- Unmanned Aerial Vehicle (UAV)-borne visual sensors offer high-resolution aerial data for diverse applications.
- Semantic perception from UAV imagery faces challenges like dense small objects, scale variations, and computational constraints.
Purpose of the Study:
- To develop a lightweight semantic perception framework for UAV visual data.
- To address challenges in aerial image analysis, including small objects and complex structures.
- To achieve a favorable accuracy-efficiency trade-off for UAV-based scene understanding.
Main Methods:
- Proposed MEKD-UAVSeg framework utilizing conflict-suppressed heterogeneous expert distillation.
- Employed a Transformer-based semantic expert and a Mamba-based spatial expert during training.
- Incorporated UAV-aware density and hard-region priors, along with a reliability routing strategy.
Main Results:
- The final inference model is a compact CNN-based segmentation network.
- MEKD-UAVSeg demonstrated a competitive accuracy-efficiency trade-off on UAVid and UDD6 datasets.
- Achieved superior performance compared to existing CNN, Transformer, Mamba, and hybrid methods.
Conclusions:
- MEKD-UAVSeg offers an effective solution for semantic perception in UAV visual data.
- The framework successfully distills knowledge from heterogeneous experts without increasing inference complexity.
- Provides a robust and efficient approach for low-altitude scene understanding and monitoring.
