基于形状特征的物体匹配方法研究  

Object matching using shape features

作　　者：彭绍武[1,2] 林倞[2] 刘乐元[1,2] 桑农[1] 

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]莲花山计算机视觉和信息科学研究院,湖北鄂州436000

出　　处：《华中科技大学学报（自然科学版）》2009年第3期58-61,共4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(2007AA01Z166);国家自然科学基金资助项目(60672162)

摘　　要：针对基于形状特征的物体匹配,对精尺度图像提出一种分层图匹配算法,采用两条马尔可夫链同时对图形进行分割与匹配采样.该方法无需额外的学习训练过程,实现了将物体对象与复杂背景分离,并同时完成了匹配参数及能量计算.对粗尺度图像提出基于方向梯度直方图算子的图像匹配算法,利用物体的整体形状信息进行全局匹配,能忽略局部纹理的干扰,并具有运算速度快的优点,且与分层图匹配算法在计算尺度图像及速度上形成互补.结果表明,分层图匹配算法在100次迭代内便能实现精确匹配.For fine scale images, a deformable layered graph matching algorithm is proposed. A sampling process iterates between two Markov chain Monte Carlo （MCMC） dynamics, which deals with the partition and matching simultaneously. It separates the target object from cluttered background, together with the matching resultand energy exported, and no extra training process is involved. For the coarse scale images, the histograms of oriented gradients （HOG） descriptor is used to perform a nearest neighbor image matching, which neglects the disturbance of local texture, and complements the deformable layer graph matching method in both scale and computing speed. Experiments show that within 100 iterations, the graph match algorithm can achieve its best result.

关 键 词：图像处理 图形匹配 吉伯斯采样 方向梯度直方图 物体检索 马尔可夫链蒙特卡罗框架 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]