基于中心点回归的大场景SAR图像舰船检测方法

Ship detection in large scene SAR images based on target center point regression

合成孔径雷达(synthetic aperture radar,SAR)图像舰船目标检测在军事和民用领域有着重要的应用.然而随着SAR图像成像能力的提升,SAR成像场景越来越大,舰船目标检测存在两个难点:一是舰船目标在整幅图像中所占比例极小,很难与周围背景分开;二是靠岸舰船目标通常密集排列,目标之间难以区分.目前常用基于锚框的检测方法容易造成大场景SAR图像中舰船目标的漏检.为解决上述问题,本文提出了基于目标中心点的大场景SAR图像舰船目标检测方法.在进行海陆快速分割的基础上,采用CenterNet无锚框检测器,通过关键点估计来定位目标的中心点,并由中心点的信息回归得到目标边界来实现目标检测,从而有效避免了基于锚框的检测方法可能存在的漏检问题.基于公开数据集SAR-ship-Dataset的实验表明,本文方法能够精确检测大场景SAR图像中的舰船目标,检测率达到92.4%;针对密集排列目标,相较于SSD、YOLO、Fast R-CNN等方法,本文方法也能够获取最优检测性能.

Ship target detection in SAR images has important applications in military and civilian fields.However, with the improvement of SAR image imaging capabilities, SAR imaging scenes are getting larger and larger,and there are two difficulties in ship target detection: first, ship targets account for a very small proportion of the entire image, and they are difficult to be separated from the surrounding background;second, the targets of docked ships are usually densely arranged, and it is difficult to distinguish among targets. Currently, commonly used anchor box based detection methods are likely to cause missed detection of ship targets in SAR images of large scenes. In order to solve the above problems, this paper proposes a ship detection method based on the target center point in the large scene SAR images. Based on the fast segmentation of land and sea, the anchor free detector CenterNet is used to locate the center point of the target through key point estimation, and the target boundary is obtained from the center point information regression to achieve target detection, thus effectively avoiding the problem of missing detection based on the anchor frame detection method. Tests based on the public dataset SAR-ship-Dataset show that the method can accurately detect ship targets in SAR images of large scenes, with a detection rate of 92.4%;for densely arranged targets, compared with SSD, YOLO, Fast-RCNN, etc, the method in this paper can also obtain the optimal detection performance.