基于多尺度多特征视觉显著性的海面舰船检测

发布时间：2018-04-16 22:36

本文选题：目标检测 + 舰船检测　；参考：《光学精密工程》2017年09期

【摘要】：为了精确地检测到舰船目标,提出了一种基于多特征、多尺度视觉显著性的海面舰船目标检测方法。该方法首先利用多尺度自适应的顶帽算法抑制云层、油污的干扰,然后提取双颜色空间特征以及边缘特征构成双四元数图像进行舰船显著性检测。由于充分利用了双四元数图像,故可对多个特征尺度进行处理,并保证不同尺度特征之间关联性。该方法还利用人眼对不同用大小的图像关注目标不同的特点对图像进行上下采样以避免漏检和检测重叠。在得到显著图后利用自适应图像分割(OTSU)算法确定舰船所在的区域,并在原图上标定、提取舰船目标。在多种海面情况下进行了实验分析,结果表明:该算法可以排除多种干扰,精确地检测到舰船目标,真正率达97.73%,虚警率低至3.37%,相较于他频域显著性检测算法在舰船检测方面有明显的优势。
[Abstract]:In order to accurately detect ship targets, a multi-feature and multi-scale visual salience method is proposed for ship target detection on the sea surface.Firstly, the multi-scale adaptive top cap algorithm is used to suppress the interference of cloud and oil pollution, and then the binary quaternion images are extracted to detect ship salience.Because the binary quaternion images are fully utilized, many feature scales can be processed and the correlation between different scale features can be ensured.The method also makes use of the characteristics of human eyes to focus on different objects of different sizes to sample the images up and down in order to avoid missing detection and detection overlap.After getting salient images, adaptive image segmentation algorithm (OTSUA) is used to determine the region where the ship is located and calibrated on the original image to extract the ship target.The experimental results show that the algorithm can eliminate many kinds of disturbances and detect ship targets accurately.The real rate is 97.73 and the false alarm rate is as low as 3.37. Compared with his significant detection algorithm in frequency domain, it has obvious advantages in ship detection.
【作者单位】：中国科学院长春光学精密机械与物理研究所中科院航空光学成像与测量重点实验室;中国科学院大学;
【基金】：吉林省重大科技攻关专项基金资助项目(No.11ZDGG001) 吉林省自然科学基金资助项目(No.20150101017JC)
【分类号】：TP751

【相似文献】