【作者】 吕杰；

【导师】 张桂林；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2005， 硕士

【摘要】 计算机视觉系统的目标是解释已有的“视觉”数据,并使用这些解释去完成任务。出于机器人工件抓取任务、自动导航、自动检测、装配任务与医学图像的分析等等大量现实应用的需要,如今目标识别已成为一个很活跃的研究领域。现有的很多目标识别方法主要是基于模板匹配和几何形状特征点提取分类等方法,这些识别方法存在很多缺点。比如基于图像匹配的方法直接比较两幅图像相应象素灰度级之间相似性的总和,这种方法并不具备方向、尺度、几何畸变等不变性,这就需要目标图像和样本图像之间的相关性很强。采用合适的目标识别方法来尽量克服目标和背景产生的畸变,这是目标识别领域的一个难点所在。特征抽取是模式识别中的关键问题之一。原始特征的数量可能很大,样本可能是处于一个高维空间中,通过映射或变换等方法可以用低维空间来表示样本。我们的目的就是为了能够在低维空间中更好地进行分类识别。本文主要研究在目标和背景在发生姿态变化﹑噪声干扰﹑部分遮挡﹑光照变化等不同畸变情况下,尽可能地从目标样本中抽取有益于识别的信息特征,提高分类识别率同时加快识别速度。本文首先深入研究了特征空间的原理,并对识别过程中的特征表达和相似性度量这两个关键步骤做了深入的研究工作和实验,实现了基于特征空间的目标识别系统,其中采用了几种有效的改进方法,有效地解决了识别过程中存在的诸如光照以及3D目标姿态估计等问题,提高了识别的鲁棒性。完成子空间目标识别算法以后,为了有效地抽取模式的分类信息,继续研究了基于Fisher 鉴别准则的线性鉴别分析方法。Fisher 最佳鉴别矢量方法的基本思想是将原来高维的模式样本投影到最佳鉴别矢量空间中以达到减少特征空间维数的目的。投影后的模式样本在新的子空间中有最大的类间距离和最小的类内距离,即模式在该空间中有最佳的可分离性。对于高维模式识别问题,原来的Fisher 鉴别准则具有一定的缺陷,会经常碰到“维数灾难”的问题。最后,本文讨论了基于流形学习的局部线性嵌入方法,它是针对非线性数据的一更多还原

【Abstract】 The aim of Computer Visual System is the interpretation for the existing "Visual" data, which can be utilized to complete tasks. Nowadays, object recognition has become an active research field, in need of a wide variety of practical applications, such as robots’ work piece snatching at tasks, auto-navigation, auto-detection, tasks equipment and analysis of medical image, and so on. Presently many methods of object recognition are mainly based on the techniques of the template matching and the classification of the pick-up for the characteristic points of the geometric shape, which possess a large number of disadvantages. For instance, a kind of method based on image matching directly compares the sum of gray of corresponding pixel in two images. But this discriminating method is not provided with the inflexibility of direction, scale and geometric aberrance, so it demands that object image and the sample image must be of strong relativity. Adopting the appropriate method to get over the aberrances of the target and background is one of the difficulties to the object recognition field. The feature extraction is one of the key problems of pattern recognition. Maybe the amount of original features is very large, sample objects maybe situate in high dimensional space, but they can be expressed in low dimensional space by mapping or transforming method or etc. We aim at classifying and recognizing objects much better in low dimensional space. This thesis mainly researches the methods of improving recognition ratio and quickening recognition speed by extracting feature information benefited to recognition from sample objects as much as possible, while the objects and backgrounds are under the aberrant conditions such as having pose variety, noise disturbance, partial occlusion, illumination alteration and etc. Firstly, this thesis lucubrates in the theory of eigenspace and does a lot of experiments and further researches in the two steps, character expressing and measuring of similarity. The object recognition system based on eigenspace was fulfilled, which adopts some effective and ameliorative means to solve the problems existing in the course of recognition such as illumination and pose estimation, and finally enhances the robustness of recognition. After finishing arithmetic of the subspace object recognition, in order to effectively extracting classifying imformation of different patterns, more research work on linear discriminant analysis methods based on Fisher discriminant criterion was done. The basis of Fisher optimal discriminant vectors method is projecting high dimensional pattern samples over optimal discriminant vectors space so as to reducing the dimensions of feature subspaces. After projecting they have maximum between-class distance and minimum within-class distance, that is, patterns have the best separability in these new feature subspaces. While coming down to high dimensional recognition, Fisher discriminant criterion has its limitation, it often encounts “Course Dimensionality”problem. At last, this thesis discusses locally linear embedding algorithm based on manifold learning method. This algorithm is a new reducing dimension method mainly focus on non-linear data. It can both reducing dimensions and maintain the topology structure of original data. As a new reducing dimension method, it has the merit of processing both non-linear and linear data, and is becoming a research hotspot in reducing dimensions of non-linear data, clustering and image division area ,etc. This thesis combines locally linear embedding method with Fisher linear discriminant analysis method to classify and recognize objects. Experiment results demonstrate that the combination method is a new effective recognition method having the merits of quick recognition speed and high recognition ratio.更多还原