Cross-Modality Object-Level Knowledge Distillation for Enhanced Underwater Sonar Object Detection

Underwater object detection plays an important role in marine safety and defense, particularly in identifying explosive-like or hazardous materials lying on the seabed. This study proposes a knowledge distillation (KD) framework to enhance the underwater object detection performance of sonar systems by leveraging an additional sensing modality. High-resolution camera images act as the teacher modality, whereas sonar images serve as the student modality. The proposed method was designed to handle heterogeneous modalities with fundamentally different structures, resolutions, and geometric representations. A YOLOX-based detection architecture equipped with a Cross-Model Object-Level Distillation (CMOLD) module was employed to transfer the object-level knowledge from the camera to the sonar. A key motivation of this study is that sonar-based detection alone often struggles with low-contrast, ambiguous, or noisy underwater returns. Although cameras may fail in deep, dark, or turbid environments, the availability of paired camera–sonar data during training enables the sonar model to learn richer high-level feature representations that cameras naturally provide. As a result, the distilled sonar model becomes more capable of detecting targets in real-world conditions where no camera input is available at inference time. This capability is particularly valuable for defense and security applications, where Autonomous Underwater Vehicles (AUVs) must operate reliably in visually degraded environments. The experimental results demonstrate that the proposed model improves the performance of the sonar compared to conventional sonar-only baselines and sonar-based KD methods. These findings highlight the potential of multimodal knowledge distillation to enhance the operational performance of underwater sensing systems and provide a practical solution for real-time autonomous underwater applications.