LMCNet: Lightweight Modality Compensation Network Via Knowledge Distillation for Salient Ship Detection Under Missing-Modality Conditions
Weibao Xue, Jiaqiu Ai, Yanan Zhu, Xinyu Sun, Yong Zhang, Gui Gao
Abstract
Salient ship detection is critical for maritime applications that require accurate localization. Although the fusion of Synthetic Aperture Radar (SAR) and Automatic Identification System (AIS) data has proven effective in enhancing saliency and suppressing background clutter, practical deployment faces two major limitations: the limited unavailability of AIS data and the high computational overhead of existing multimodal models. These limitations pose a fundamental challenge in balancing detection accuracy and efficiency under missing modality and resource-constrained environments. To address this issue, a Lightweight Modality Compensation Network (LMCNet) is proposed. A multimodal teacher network is trained with SAR and AIS inputs to learn rich and complementary representations. Meanwhile, a compact, single-modality student network that relies only on SAR is designed to support low-cost, real-time deployment. To enable robust knowledge compensation and transfer, this paper designs a knowledge distillation strategy consisting of three modules: structure-aware attention distillation for spatial alignment, cross-head teacher distillation for semantic enhancement, and adaptive loss scheduling for dynamic optimization. This unified design allows the student model to inherit spatial precision and semantic awareness from the teacher, achieving strong performance even with limited input modalities. Extensive experiments on two datasets show that our distilled student model improves <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">E<sub>ξ</sub></i> E by 1.39% and F<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">w</sup><italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><sub>β</sub></i> by 4.37% compared to state-of-the-art methods, demonstrating its superior balance of accuracy and efficiency in real-world deployment scenarios.