Lightweight Underwater Sonar Object Detection via RGB-Guided Heterogeneous Distillation.

Underwater object detection is a fundamental task in underwater sensing and is generally approached using either optical or sonar sensors. Although optical imaging provides rich semantic information, it is highly susceptible to water turbidity and illumination variations. By contrast, sonar imaging can effectively overcome visibility limitations, yet it suffers from severe speckle noise and blurred object contours. Moreover, resource-limited platforms impose strict demands on model lightweightness and real-time performance. To this end, this paper proposes a novel cross-modal heterogeneous distillation method (CMHD) to balance detection accuracy and computational complexity. CMHD performs cross-modal knowledge transfer by leveraging the rich semantics of RGB images to enhance sonar feature representation, compensating for the information deficiency of the sonar modality. Meanwhile, a heterogeneous distillation scheme compresses the detection capability of a high-capacity teacher YOLOX-M into a lightweight student YOLOX-S-Ghost, enabling strong feature extraction under a highly compact model. To mitigate the modality gap and geometric inconsistency between RGB and sonar modalities, we design a branch-aware heterogeneous distillation strategy. To improve detection accuracy and reduce model parameters, the student network incorporates Coordinate Attention (CA) in its backbone and adopts a lightweight neck design. Experiments on the UXO† dataset demonstrate that CMHD achieves 79.6% mAP and 82.6% mAR, significantly outperforming the compared representative methods and serving as an accurate, efficient, and lightweight solution for underwater sonar object detection.