基于视觉显著性的无监督海面舰船检测与识别

Detection and identification of unsupervised ships and warships on sea surface based on visual saliency

在航天航空光学遥感舰船目标检测中,受大气、光照、云雾和海岛等海面不确定条件的影响,传统的舰船检测方法存在检测效率低和可靠性差等问题,因此,本文提出一种无监督海面舰船目标自动检测方法。该方法以视觉显著性为依据,结合多显著性检测模型快速搜索海面目标,生成显著图后对其进行粗分割,对提取的目标切片做标记并进行精细分割,利用改进的Hough变换旋转目标主轴以保证目标对Y轴的对称性;对可能检测到的厚重云层和岛屿等伪目标使用梯度方向特征进行鉴别,通过判定目标在360°范围内8个区间的梯度幅度统计值,确认舰船目标及去除伪目标。实验结果表明,该舰船检测方法能够成功提取海面上大小不同,位置随机分布的舰船目标,准确获取舰船目标的数量和位置信息,在大量真实光学遥感图像上的测试结果显示,本文方法检测准确率高于93%,通过目标鉴别处理,剔除伪目标后,虚警率可低于4%,鲁棒性较强。

In target detection on aerospace optical remote sensing,due to the interference of uncertain conditions on sea surface such as atmosphere, solar radiation, cloud and mist, islands and others, traditional ship detection methods always have some defects such as low detection efficiency, poor reliability. Therefore, the author proposed an unsupervised ship automatic detection method. In this method, visual saliency was combined with multi-saliency detection model for fast searching of sea- surface targets~ after saliency map was generated, a rough segmentation was conducted on it, then extracted target slice was marked and fine segmentation was implemented, subsequently, improved Hough transformation was used to rotate principal axis of target for ensuring the symmetry of targets to Y axis; the characteristics of gradient direction was applied to recognize phony targets such as thick clouds layer, islands and others that may be detected, the gradient and amplitude statistical value of those targets in 8 intervals on all directions were judged to identify target ships and warships and eliminate phony targets. The experimental result indicates that the detection method of ships and warships can be used to successfully extract target ships and warships which are in different size and random distributed on sea surface for obtaining accurate quantity and location information about them. In test on a large number of authentic optical remote sensing pictures, the detection accuracy rate of proposed method is higher than 93%, while the false alarm rate is lower than 4% through target identification and treatment and phony target elimination, which has strong robustness.