灰度不均匀的曲线形态检测方法及应用研究-控制科学与工程专业论文.docx

华中科技大学博士学位论文 华 中 科 技 大 学 博 士 学 位 论 文 II万方数据 II 万方数据 再次，宽线是特殊的曲线形态，在图像中表现为条状的区域。受噪声和光照等 影响，宽线检测易出现断裂甚至漏检的问题。本文利用梯度方向的延续性，提出了 一种有效的宽线检测算法。该算法提出线增强算子以增强线特征，并提出接近方向 和采用假设验证法得到线候选目标，再利用背景灰度具有相似性的特点，来确定线 是否为宽线，从而排除杂乱线条的干扰，接着沿着线的剖面检测确定另一侧宽线， 并利用拐点处具有无延续性的特点检测拐点。实验结果表明，该算法能有效地检测 变对比度的宽线。 然后，在轮廓提取的基础上，轮廓的描述与检测是目标检测的最重要的课题之 一。由于受视角、光照和噪声等影响，轮廓可能存在尺度、旋转、毛刺、微小缺失 等问题。针对这些问题，本文提出了基于改进的曲率函数和行程编码的局部轮廓描 述子和基于空间对称性的全局轮廓描述子，这两个轮廓描述子对旋转和尺度具有不 变性，且对微小缺失和毛刺等具有鲁棒性。最后，本文利用隐马尔可夫模型 (HMM) 框架，提出了基于 HMM 模型的轮廓描述子匹配算法，并对 MPEG-7 Shape CE Part B 和 Kimia 99 数据库进行测试，实验表明，本文提出的匹配算法容错能力强，适合 轮廓描述子序列的匹配。 最后，针对货车故障定位与检测的应用，从大尺度曲线形态故障部位定位和感 兴趣区域典型曲线形态检测两方面，系统地研究了线、圆以及宽线等各形态的故障 部位定位与检测方法，研制和开发了货车故障定位与检测应用软件系统，并已投入 铁路货车故障检测的实际应用。两年四个月，现场采集的七十一万节货车车厢的图 像数据样本的测试表明，本文提出的货车故障定位与检测算法的准确率至少达到 90%，满足铁道部对货车故障检测系统的要求。 关键词：灰度不均匀 曲线检测 宽线检测 轮廓描述 序列匹配 I万方数据 I 万方数据 Abstract Curve shape is the important object feature. It includes line, circle, curve, wide line, contour and so on. It has been widely used in object detection such as characters, manmade object, and classical building detection. We call the image intensity inhomogeneity image if the image has great noise and gray is inhomogeneous. intensity inhomogeneity leads to cracking or false edge detection, false or missing curve detection, and incomplete contour. It is still a task of great challenge to improve the efficiency and reliability of intensity inhomogeneity curve shape detection. Consider the truck fault location and detection system as the application of curve shape detection, curve shape detection is mainly divided into four parts. First of all, the watershed algorithm based on structure is studied to obtain the closed edge or contour. Second, curve and wide line detection is studied by perceptual organization. Third, contour descriptor and its matching algorithm which are roubust to scale and rotation are studied. At last, truck fault location and detection system is introduced, and methods of curve detection of truck image are designed. The main work and