基于流形学习和流形高阶近似的图像距离度量  

Image distance measurement based on manifold learning and manifold high-order approximation

作　　者：周华[1] 蔡超[1] 丁明跃[2] 

机构地区：[1]华中科技大学图像识别与人工智能研究所,湖北武汉430074 [2]华中科技大学生命科学与技术学院,湖北武汉430074

出　　处：《华中科技大学学报（自然科学版）》2012年第3期1-4,17,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家高技术研究发展计划资助项目(2007AA12Z166)

摘　　要：为克服切距离方法的不足,提出了流形高阶近似距离(HMD).HMD度量方法通过最大差异延展方法(MVU)学习出非线性图像流形的内蕴变量,然后在原型图像处用基于流形高阶泰勒展开式的非线性曲面来局部近似图像流形.HMD定义为待识别图像与图像流形的高阶近似曲面间的最小距离,通过计算待识别图像与图像流形上多个基准图像间的HMD距离能够实现图像的分类和识别.人脸识别和手写数字识别的结果表明:HMD距离在识别精度和稳定性上要优于切距离和当前一些典型的图像距离度量方法.To overcome the shortcomings of tangent distance,an image distance high-order approximated manifold distance（HMD） was proposed.In HMD metric intrinsic variables of nonlinear image manifold were revealed by maximum variance unfolding（MVU） and image manifold was locally approximated at prototype by curved surface based on high-order Taylor expansions with respect to intrinsic variables.HMD was defined as the minimum distance between the query image and the approximated curved surface of image manifold.The image classification and recognition can be achieved by computing HMD distances between the query image and the landmarks of image manifold.A series of face recognition and handwritten digit recognition experiments demonstrate that HMD can not only achieve higher recognition accuracy but also has more stability of classification than several state-of-the-art distance metrics.

关 键 词：图像距离 流形学习 流形近似 最大差异延展 切距离 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]