【作者】 张浩；

【导师】 周泽魁； 侯迪波；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2008， 博士

【摘要】 多源信息融合技术是协同利用多源信息以获得对同一事物或目标更客观、更本质认识的信息综合处理技术,最早出现于20世纪70年代。从诞生起,多源信息融合技术就吸引了人们越来越多的关注,对整个信息科学产生了极大的影响,图像融合是多源信息融合技术主要应用领域之一。研究表明,在人类接受到的信息中图像等视觉信息所占比重达到75%,图像中蕴涵着丰富的信息,如何获取这些信息并采用有效的手段进行综合利用是图像模式识别领域中的重要课题。本文利用多源信息融合思想,全文围绕如何融合多图像信息以实现图像边缘特征稳定提取,及利用边缘及其二次特征进行图像配准为主线展开研究。其中图像边缘检测与图像配准是文中两个联系紧密的主要部分。图像边缘是一种重要的视觉信息,边缘检测是计算机视觉后续工作的基础和前提。图像配准属于图像模式识别研究领域,主要用于多图像融合,在遥感和军事领域有着重大的现实意义。本文对边缘检测、边缘二次特征提取及图像配准做了一定研究,以提高视觉检测、目标识别及图像合成的稳定性和实时性为研究目标。并将研究应用于AGVS系统、SAR图像处理及数字字符识别等领域。本文的主要工作和创新点如下:(1)提出一种改进鲁棒Fisher降维方法,通过引入权值函数和调整因子降低“外点”对降维方向计算的干扰,对多维特征向量进行降维,克服了应用贝叶斯理论进行统计推理时,特征维数与小训练样本、特征相关性及计算效率的矛盾。根据边缘的本质特点,构造多级尺度上的绝对梯队和相对梯度作为多维图像信息,降维后利用贝叶斯理论加以融合完成边缘检测。并将该算法运用于复杂场景边缘检测及AGVS系统导引线检测中。(2)对DS证据理论进行研究,提出一种改进相关证据合成方法,利用证据可信度分布计算出证据间相关度,克服了相关证据合成时的超估计及提高合成结果的合理性,并将传统两相关证据合成推广至多相关证据合成。进而建立了一种基于证据理论的非学习直接融合边缘检测模型,将该模型应用于SAR图像边缘检测中,融合了两级尺度上的ROA算子及梯度算子。该方法最大优点是无需学习过程,可对多图像信息进行直接融合实现边缘检测。(3)提出一种基于边缘拟合的图像配准方法。以边缘为基本特征,通过可变精度拟合算法将边缘拟合成方便表达的直线特征。根据边缘直线与图像配准的关系,对其进行筛选和可信度赋值,并利用加权投票算法实现图像配准,保证配准的稳定性和实时性。(4)根据图像配准特点,将Hausdorff距离转换成归一化的相似度值。改进边缘曲率计算,更稳定地提取边缘角点。并利用“Local jet”、边缘及距离约束,提高计算效率。利用改进Hausdorff距离和形态学实现快速稳定的图像配准。将上述方法应用于AGVS系统路标识别、SAR图像和航拍图像配准中,结果表明该方法有效可行。(5)图像匹配是图像配准的一种特殊形式,针对钢坯标号和AGVS系统工位号识别,分别提出两种数字字符识别方法。针对钢坯标号,利用隶属度概念改进传统模板匹配方法,对字符和模板进行隶属度分配,提高识别稳定性。对于AGVS系统工位号,则提出一次倾斜校正和二次自校正,使失真字符接近正常字符,并利用两级模板匹配方法对字符进行快速稳定识别。更多还原

【Abstract】 Multi-source information fusion is the technique that utilizes different information from the same object synthetically in order to obtain the more objective and entitative information for this object. Multi-source information fusion technique had its beginnings the early 1970’s firstly. Ever since its emergence, this technique has made a strong impact on entire information science, such as image fusion. At the same time, studies show that vision information achieves 75% in all the information received by human. Images contain rich information, and how to obtain this information and utilize it synthetically by effective method is important subject in the pattern recognition field.This dissertation applies the idea of multi-source information fusion. The whole dissertation is closely linked to edge detection based on fusing multiple information, secondary edge feature extraction and register multiple images based on edge feature. The edge detection and image registration are the two main topics in this dissertation. Image edge is the most important information in image, and edge detection is the basis for the follow-up work generally. And image registration belongs to image pattern recognition research field. And the image registration technique is used mainly for multiple images fusion and has practical significance in the field of military and remote sensing.The work in the dissertation is focus on the studying of edge detection, secondary edge feature extraction and image registration. The main goal is to improve the stability and real-time performance. And this study is applied to AGVS, SAR image and numeral character recognition, etc. The major contributions in the dissertation are as follows:(1) To overcome the contradiction between multi-dimensional vector and small training set, element correlation in vectors, computational efficiency when applying Bayesian statistical inference theory, an improved dimensional reduction method is proposed for multi-dimensional data space. The reduction direction computation is robust to outliers by introducing weight function and adjustment factor. Absolute gradient and relative gradient at several scales are constructed to be the multi-dimensional image information. Firstly, the multi-dimensional image information is reduced. Then Bayesian statistical inference theory is employed to fuse multiple features to accomplish edge detection. This method is applied to edge detection in complex scene and guided line detection in AGVS.(2) With the research on DS evidence theory, an improved relevant evidence fusion algorithm is proposed. This method overcomes overestimation with evidence fusion, improve the reasonableness of fusion result and suitable for multiple relevant evidence fusion. Correlation degree is computed by the contribution of evidence reliability. Then a novel edge detection model based on evidence theory is also constructed. This edge detection method is used to fuse ROA operator responses and gradient operator responses at two scales in order to realize SAR image edge detection. The major advantage of this method is fusing information directly to realize edge detection without learning process.(3) A novel image registration method based on edge fitting is proposed. This method is based on edge feature and fit these edges to straight line feature which can be expressed with mathematical method easily through an alterable precision fitting algorithm. Based the relationship of edge and image registration, straight line is screened and set a reliability value. And a weighted voting algorithm is employed to perform registration process to ensure the registration performance.(4) According to the character of image registration, traditional Hasudorff distance is transformed to a normalized similarity value.. The traditional edge curvature computation is adjusted in order to extract the edge corners more stable. "Local jet", edge geometry and distance sequence restriction are used to improve the computation efficiency. Improved Hausdorff distance and morphological method are employed to perform registration process. This registration method is applied to route mark recognition in AGVS, SAR and aerial images registration. Experiments show that this registration method is effective and feasible.(5) Image template matching is the spectral form of image registration. For the two application background of characters on steel billets and working station mark in AGVS, two novel numeral character recognition methods are proposed. For characters on steel billets, classical template matching technique was improved by membership assignment conception in order to improve the recognition stability. For working station mark in AGVS, first skew rectify and second own rectify are proposed to make distorted character closed to natural character. Then the two stage template matching method is employed to recognize the numeral guide character fast and stably.更多还原