WS-Det: an efficient whale shark detection via integrated feature modulation module

Abstract

Whale sharks are vital plankton feeders that help maintain marine ecosystem stability and serve as key ecotourism assets supporting Gorontalo's blue economy. Reliable detection and identification systems are essential for monitoring their habitats and behaviors. Vision-based methods offer non-invasive observation without physical contact or electromagnetic interference. YOLOv12, a state-of-the-art object detector, achieves high localization accuracy with its lightweight "nano" variant. However, lightweight models often struggle with complex feature discrimination, which can be mitigated by integrating enhancement blocks to boost detection performance while maintaining computational efficiency. In this paper, we introduce a novel vision-based whale shark detection framework (WS-Det), which enhances YOLOv12-nano with MOGA attention modules. The proposed model incorporates lightweight convolutional operations to ensure efficient feature extraction while keeping computational cost low. The Integrated Feature Modulation (IFM) module improves backbone performance by selectively emphasizing spatially salient regions while interacting with multi-spatial-frequency information, thereby improving the overall robustness of feature learning. To support this study, we also present a dedicated whale shark dataset comprising underwater and aerial (drone-based) images for direct and practical monitoring. Experimental evaluations demonstrate that the WS-Det outperforms the baseline YOLOv12-nano and other efficient detectors, achieving higher detection precision. On the proposed dataset, the model achieves an [email protected] of 0.979 and an [email protected]:0.95 of 0.868. It contains 2,191,459 parameters and requires 6.3 GFLOPs, indicating a low computational cost. Inference speed evaluations show the model runs at 17.19 FPS on a desktop PC and 6.64 FPS on a Raspberry Pi 5, confirming its suitability for deployment on resource-constrained devices.