IGC-Net: Integrating gated mechanism and complex-valued convolutions network for overwater object detection

In real-world overwater scenarios, detecting occluded or distant objects is common challenges. In this paper, we initially construct a novel dataset SeaShips24790 for evaluating the performance of overwater object detectors, which includes 24,790 diverse overwater object annotations, especially focusing on small-scale objects. Subsequently, a new deep-learning network that integrates gated mechanism and complex-valued convolutions, termed IGC-Net, is proposed to tackle the challenges of object occlusion and small object detection in overwater scenarios. It employs the gating mechanism to selectively enhance or suppress features and incorporates complex-valued modules, including complex-valued convolutions, for fusing multi-scale feature maps. Additionally, a two-stage multi-scale feature fusion is used, comprising pre-fusion and post-fusion stages. Experimental results demonstrate that our proposed IGC-Net achieves state-of-the-art (SOTA) performance across several overwater object detection datasets. The SeaShips24790 dataset will be made available as requested.