Related Experiment Video
Updated: Aug 5, 2026
09:32
Development of New Methods for Quantifying Fish Density Using Underwater Stereo-video Tools
Published on: November 20, 2017
Lightweight Underwater Sonar Object Detection via RGB-Guided Heterogeneous Distillation
Qianqian Qiao1, Jia Liu2, Feng Liu3
1State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210093, China.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a novel cross-modal heterogeneous distillation method (CMHD) for underwater object detection. CMHD enhances sonar data using RGB images and compresses a large model into a lightweight one for efficient, accurate sonar detection.
Area of Science:
- Robotics and Autonomous Systems
- Computer Vision
- Marine Technology
Background:
- Underwater object detection faces challenges with optical sensors (turbidity, illumination) and sonar sensors (speckle noise, blurred contours).
- Resource-limited platforms require lightweight and real-time underwater detection models.
- Existing methods struggle to balance accuracy and computational efficiency in diverse underwater conditions.
Purpose of the Study:
- To propose a novel cross-modal heterogeneous distillation method (CMHD) for accurate and efficient underwater object detection.
- To leverage RGB image semantics to enhance sonar feature representation and compensate for sonar data deficiencies.
- To develop a lightweight yet powerful detection model for resource-constrained underwater platforms.
Main Methods:
- Implemented a cross-modal heterogeneous distillation approach (CMHD) for knowledge transfer between RGB and sonar modalities.
- Utilized a teacher-student model compression strategy (YOLOX-M to YOLOX-S-Ghost) for a lightweight student network.
- Introduced a branch-aware heterogeneous distillation strategy to address modality gaps and geometric inconsistencies.
- Incorporated Coordinate Attention (CA) and a lightweight neck in the student network's backbone for improved performance.
Main Results:
- The proposed CMHD method achieved 79.6% mean Average Precision (mAP) and 82.6% mean Average Recall (mAR) on the UXO dataset.
- CMHD significantly outperformed existing representative methods in underwater sonar object detection.
- The method demonstrated a balance between detection accuracy and computational complexity, offering a lightweight solution.
Conclusions:
- CMHD provides an accurate, efficient, and lightweight solution for underwater sonar object detection.
- Cross-modal knowledge distillation is effective in overcoming limitations of individual sensor modalities.
- The developed model is suitable for deployment on resource-limited underwater sensing platforms.
