光学 精密工程, 2018, 26 (3): 723, 网络出版: 2018-04-25   

光学遥感图像中复杂海背景下的舰船检测

Ship detection of complex sea background in optical remote sensing images
作者单位
1 中国科学院 长春光学精密机械与物理研究所, 吉林 长春 130033
2 中国科学院大学, 北京 100039
摘要
本文针对光学遥感图像中复杂海背景下的舰船检测问题, 提出一种快速精确的舰船检测方法。首先, 基于最大对称环绕显著性检测完成初始目标候选区域提取, 并结合一种基于元胞自动机的同步更新机制, 利用图像局部相似性和舰船目标几何特征, 对初始目标候选区域进行更新, 并通过OTSU算法获取最终目标候选区域; 然后, 根据舰船目标的固有特性对方向梯度直方图特征进行改进, 提出一种新的表征舰船特性的边缘-方向梯度直方图特征对舰船目标进行描述, 与传统HOG特征相比, 这种特征向量侧重于对边缘特征的描述, 对梯度向量鲁棒性更强, 并且仅为一个24维的特征向量, 计算复杂度低; 最后, 通过构建的训练库完成AdaBoost分类器的训练, 并利用训练完成后的AdaBoost分类器完成目标的最终判别确认。本文的检测算法, 针对尺寸为1 024 pixel×1 024 pixel的遥感图像, 检测时间为2.386 0 s, 召回率为97.4%, 检测精度为97.2%。实验表明, 本文提出算法的检测性能优于目前主流的舰船检测算法, 在检测时间和检测精度上都能够满足实际工程需要。
Abstract
In this paper, a fast and accurate ship target detection method wa proposed for ship detection in optical remote sensing image. The “coarse-to-fine” strategy was applied, which contains mainly three stages: the candidate regions extraction, building the candidate regions’ descriptor and the candidate regions discrimination by reducing the false alarms to confirm the real ship targets. In the first stage, first the initial saliency map was extracted by the maximum symmetric surround method, which was based on the visual attention mechanism, and updated according to the local similarity via a updating mechanism of cellular automata; then, the final saliency map was segmented by OTSU algorithm to obtain binary image; finally salient regions were extracted from the segmented binary image, and filtered roughly by the ship objectives' geometric features. In the second stage, a new descriptor, named edge-histogram of oriented gradient (E-HOG), was proposed to describe the ship target. The E-HOG feature was an improvement of the traditional HOG feature, based on the inherent characteristics of the ship targets. Compared to the traditional HOG feature, the E-HOG feature limited the statistical scale into the edge of the salient regions, for the purpose of reducing the influence of the variability of oriented gradient, and reducing computation complexity. On one hand, the descriptor could discriminate the ship objectives from others like cloud, islands and wave; on the other hand, the descriptor was insensitive to the size of the ship objectives, which reinforce the robustness of the approach. In the third stage, the AdaBoost classifier was used to confirm the real ship targets by eliminating the false alarms. We intercept 517 positive samples and 624 negative samples from the remote sensing images to train the AdaBoost classifier. The size of these training samples ranges from 20 pixel×10 pixel to 200 pixel×120 pixel, where the positive samples include different types of ship targets, and the negative samples include non-ship targets such as clouds, islands, coastlines, waves and sea floating objects. In this paper, the detection time is 2.386 0 s for the 1 024 pixel× 1 024 pixel remote sensing image, the recall rate is 97.4%, and the detection precision is 97.2%. Experiments demonstrated that the detection performance of the proposed method outperforms that of the state-of-the-art methods, and it can meet the actual engineering requirements in the detection time and detection precision.

王慧利, 朱明, 蔺春波, 陈典兵, 杨航. 光学遥感图像中复杂海背景下的舰船检测[J]. 光学 精密工程, 2018, 26(3): 723. WANG Hui-li, ZHU Ming, LIN Chun-bo, CHEN Dian-bing, YANG Hang. Ship detection of complex sea background in optical remote sensing images[J]. Optics and Precision Engineering, 2018, 26(3): 723.

本文已被 9 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!