【作者】 庄德文；

【导师】 王守觉；

【作者基本信息】 浙江工业大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 高维仿生信息几何理论借鉴了人的认知理论,它既是一种机器学习理论,也是重要的信息处理指导方法。本文主要从以下几个方面作探讨,并得出一些有益的结果。高维仿生信息几何理论和计算过程都是从高维空间的几何图形概念出发,用空间中的点集来描述子空间,借助于平面几何简单运算组合迭代实现高维空间中复杂的计算,并有了一些成功的范例,但缺乏几何学中公理化体系的优点,本文通过引入几何线性代数理论,对高维空间几何作系统描述,并形成一个自洽的公理体系。基于高维仿生信息几何理论的多权值神经元模型是仿生模式识别理论中的一个重要成果,本文在回顾人工神经网络主要成果的基础上,提出了多权值神经元模型学习的一个新算法,并成功的应用于说话人识别,通过与熟知的说话人识别算法比较,证明了所提出算法的有效性。人脸识别是机器视觉研究中的一个热点问题,本文在系统概述人脸识别算法的基础上,结合模式识别中的一个热点研究领域——核方法,提出了基于核最小球的人脸识别新算法,实验结果表明该算法的有效性。Gabor滤波在空间域和频率域给出最优分辨率,对于提取图像纹理特征尤其有利,并在人脸识别等领域获得广泛的关注,本文在概述Gabor滤波理论及在人脸识别领域应用的有关文献基础上,结合高维仿生信息几何理论,提出了基于最小生成树的仿生模式识别算法,并应用于每类仅有单个训练样本情形下的人脸识别,对于光照变化情形的人脸识别问题作了具体的研究,并与基准算法比较,实验结果验证了所提出算法的有效性。Log-Gabor滤波可以最大化空间局域化,本文提出了基于Log-Gabor滤波二值变换的人脸识别算法,在存储和计算效率上都有改进,针对JDL-A人脸库中姿势、光照、表情、饰物、远近变化条件下的人脸识别问题,所给出的算法相比基准算法都有不同程度的提高。随着计算机硬件的革新和机器学习理论的进展,基于内容的图像检索研究吸引着越来越多研究者的兴趣,并为形成了交叉学科提供平台。本文基于高维信息几何思想,分别研究了整体特征和区域特征的图像检索,给出了一个图像检索原型系统,以COREL子集数据库为实验对象,得到与SIMPLIcity系统类似或更优的检索结果;结合相关反馈技术,利用线性判别分析,对视觉特征作降维处理所得到的低维特征空间表示,可更好的反映其高层语义信息,进一步减少底层视觉特征与高层语义特征之间的差距。更多还原

【Abstract】 Based on human cognition, geometry theory of high-dimensional biomimetic information is not only a kind of machine learning theory, but also an important guidance method for information processing. Following aspects of the theory are discussed in this paper, and some useful results are obtained.Geometry theory of high-dimensional biomimetic information and its calculation start from the concept of geometrical figures, and describe the sub-space with a point set in space. Complex computation in high-dimensional space is conducted by means of simple Euclidean geometry iterative operations. Although some successful cases have been made, the advantages of axiomatic system in geometry are not involved. With the introducing the geometric linear algebra, the author systematically describes the high-dimensional geometry and establishes a self-consistent axioms system.The multi-weight neuron model based on the geometry theory of high-dimensional biomimetic information is one of the important fruits of the Biomimetic Pattern Recognition. A new algorithm about multi-weight neuron model is presented based on the review of main results of artificial neural network, and then it is applied to speaker recognition successfully. Comparing it with the well-known speaker recognition algorithm demonstrate the effectiveness of our algorithm.Face recognition is an attractive issue in the machine vision field. Based on the review of face recognition in detail, the author combines the kernel method, proposes a new algorithm based on kernel smallest ball for face recognition, the experimental results demonstrate the effectiveness of the algorithm. Gabor filter gives the best resolution in spatial domain and frequency. It is beneficial for extracting the image texture futures and concerned in face recognition area widely. The theory of Gabor filter and its relevant papers are summarized in this paper. Combined with the theory of biomimetic patter recognition, a minimum spanning tree algorithms is proposed in this paper. This algorithm is applied in the face recognition experiment in the extreme case of single training sample in each class. In addition, the face recognition in illumination changes is deeply studied in this paper, and its effectiveness is proved. Log-Gabor filter can maximize space location. A face recognition algorithm based on the Log-Gabor filter of binary transform is proposed in this paper. The storage and computing efficiency is improved. The new algorithms is more effective than the benchmark algorithms in recognizing faces with different positions, lights, expressions, accessories and distances in the JDL-A face database.With the innovation of computer hardware and the progress of machine learning, more and more researchers are attracted by the technology of content-based image retrieval. On the basis of the high-dimensional biomimetic information geometry, the author studies the image retrieval by global features and the regional features. A prototype system of image retrieval is provided. A retrieval result similar with SIMPLIcity system or better is obtained on the COREL subset database. Combining with the relevance feedback technique, the author uses liner discrimination analysis for dimensional reduction, and the low-dimensional feature obtained can better reflect the high-level semantic information and further reduce the gap between low-level vision features and high-level semantic.更多还原