【作者】 刘丽丽；

【导师】 王树国； 付宜利；

【作者基本信息】 哈尔滨工业大学 ， 机械电子工程， 2008， 硕士

【摘要】 港口是重要的军用和民用建筑,遥感图像在军事侦察、精确打击和民用方面有很多应用,因此研究遥感图像中港口目标的自动检测与识别具有重要意义,但港口目标因环境和特征的复杂性使其识别技术成为图像处理领域难点之一。本文首先检测遥感图像中是否存在港口目标,再对检测出的港口目标进行基于特征点的Hausdorff距离匹配,由匹配结果完成港口目标识别。根据港口灰度、纹理、形状等特征,分析可用于检测的港口知识。对遥感图像进行阈值分割、边缘检测和Hough变换等操作后,由港口中防波堤间距离、两侧灰度和所围区域面积判断提取的直线是否属于港口目标,检测出港口目标的存在和在实测图中的位置。为保证港口目标特征点检测的准确性,在识别时首先采用主动活动轮廓(Snake)模型提取港口目标轮廓。Snake模型可以在模型本身的内部能量和由图像产生的外部能量的共同作用使模型不断逼近目标轮廓。改进Snake内力计算方法,避免向强边缘聚集。采用梯度矢量场(Gradient Vector Flow, GVF)作为Snake外力,扩大模型的捕获范围。实验证明由Snake模型提取的港口目标轮廓具有很高的精度。采用基于特征点的Hausdorff距离匹配识别港口目标。通过分析比较各种特征点提取算法,最小吸收核同值区(Smallest Univalue Segment Assimilating Nucleus, SUSAN)角点检测算子具有计算量小、精度高、稳定性好、对噪声不敏感等优点,可以很好的描述目标轮廓特征。采用自适应方法计算灰度阈值和几何阈值;采用稳健有效的相似比较函数;通过吸收核质心与中心的距离去除虚假角点。Hausdorff距离描述两个模糊点集之间的相似度,本文对其进行了两部分改进,即部分Hausdorff距离和基于角点特征响应值的Hausdorff距离,以消除出格点和图像遮挡对匹配结果的影响。在VC++6.0上建立实验平台,利用基于内容的相似搜索中的K紧邻(K Nearest Neighbor, KNN)查询方法评价本文算法,得到回召率—检索精度曲线,根据实验结果与直线段匹配、不变矩匹配比较,证明本文算法的优越性。更多还原

【Abstract】 Harbor is an important military and civil building; remote sensing image has many applications in military reconnaissance, precision attack and civil activities, so it has great importance to detect and recognize harbor target automatically. However, because harbor environment and structure are very complex, harbor target recognition is a difficult point in image processing area. Firstly, this dissertation detected harbor target in remote sensing image, judged its existence, secondly matched the detected harbor target based on feature points, finished harbor target recognition through matching results.Based on harbor target gray, texture, shape and structure features, harbor knowledge was analyzed for detecting harbor target. After processing procedures of threshold segmentation, edge detection, hough transform, this dissertation judged these detected lines whether belonged to harbor target based on breakwaters distance, breakwaters bilateral gray and area size enclosed by breakwaters, detected the existence and location of harbor target in remote sensing image.In recognition phase, this dissertation firstly used initiative active contour model (Snake) to pick up harbor contour which ensures accuracy of harbor feature points. Snake model can approach target contour constantly under the action of model internal force and external force caused by image. This dissertation gave an improved method to compute internal force, which avoids snake model assembling to strong edges; taked gradient vector flow (GVF) as external force, which enlarges snake model’s capture area. Experiments proved that harbor target contour picked-up by snake model algorithm has great precision.Then, harbor target was recognized by hausdorff distance matching method based on feature points. By analyzing and comparing several feature points picking-up algorithms, susan(smallest univalue segment assimilating nucleus) operator has advantages of low computational complexity, high resolution, strong anti-noise and good invariability, it can describe target contour features perfectly. Gray and geometric thresholds were obtained with an adaptive method; this dissertation used robust and effective similar function; the distance between assimilating nucleus centroid and its center was calculated to remove false corner points. Hausdorff distance presents the similarity of two fuzzy point sets. Two improvements about hausdorff distance were given. They were: partial hausdorff distance and hausdorff distance based on corner feature response value, which were used to avoid influences of partial visibility and outlier.Finally, an experiment platform was builded up on VC++6.0. KNN (K nearest neighbor) quiry method in similarity search based on contents was used to evaluate this dissertation algorithm, and recall ratio—precision ratio curve was obtained. Trough comparing with linear matching and invariability matching based on the experimental results, this dissertation algorithm was proved to be superior.更多还原