CSBBNet: A Specialized Detection Method for Corner Reflector Targets via a Cross-Shaped Bounding Box Network

In synthetic aperture radar (SAR) maritime target detection tasks, corner reflector targets (CRTs) and their arrays can easily interfere with the accurate detection of ship targets, significantly increasing the misdetection rate and false alarm rate of detectors. Current deep learning-based research on SAR maritime target detection primarily focuses on ship targets, while dedicated detection methods addressing corner reflector interference have not yet established a comprehensive research framework. There remains a lack of theoretical innovation in detection principles for such targets. To address these issues, utilizing the prior knowledge of cross-shaped structures exhibited by marine CRTs in SAR images, we propose an innovative cross-shaped bounding box (CSBB) annotation strategy and design a novel dedicated detection network CSBBNet. The proposed method is constructed through three innovative component modules, namely the cross-shaped spatial feature perception (CSSFP) module, the wavelet cross-shaped attention downsampling (WCSAD) module, and the cross-shaped attention detection head (CSAD-Head). Additionally, to ensure effective training, we propose a cross-shaped intersection over union (CS-IoU) loss function. Comparative experiments with state-of-the-art methods demonstrate that our approach exhibits efficient detection capabilities for CRTs. Ablation experiment results validate the effectiveness of the proposed component architectures.