Key points are not available for this paper at this time.
Ship detection plays a crucial role in a variety of military and civilian marine inspection applications. Infrared images are irreplaceable data sources for ship detection due to their strong adaptability and excellent all-weather reconnaissance ability. However, previous researches mainly focus on visible light or synthetic aperture radar (SAR) ship detection, while infrared ship detection is left in a huge blind spot. The main obstacles to this dilemma lie in the absence of public datasets, small scale, and poor semantic information of infrared ships, and severe clutter in complex ocean environments. To address the above challenges, we propose a knowledge-driven context perception network (KCPNet) and construct a public dataset called infrared ship detection dataset (ISDD). In KCPNet, aiming at the small scale of infrared ships, a balanced feature fusion network (BFF-Net) is proposed to balance information from all backbone layers and generate nonlocal features with balanced receptive fields. Moreover, considering the key role of contextual information, a contextual attention network (CA-Net) is designed to improve robustness in complex scenes by enhancing target and contextual information and suppressing clutter. Inspired by prior knowledge of human cognitive processes, we construct a novel knowledge-driven prediction head to autonomously learn visual features and back-propagate the knowledge throughout the whole network, which can efficiently reduce false alarms. Extensive experiments demonstrate that the proposed KCPNet achieves state-of-the-art performance on ISDD. Source codes and ISDD are accessible at https://github.com/yaqihan-9898 .
Han et al. (Fri,) studied this question.