隐马尔可夫后处理模型在视频人脸识别中应用.pdfVIP

第 30 卷第 4期 计算机应用 Vo.l 30 No. 4 2010年 4月 Journal of ComputerApp lications Apr. 2010 : 100 1- 9081( 20 10) 04- 0960- 04 代 毅, 肖国强, 宋 刚 (, 400715) (h ekuh eku@ 21cn. com; gq iao@ sw u. edu. cn) : 现有的多数人脸识别系统都专注于如何提高人脸识别算法的性能, 但 乏一种对数据源(人脸样本)进 行分析和评估的机制针对此问题, 提出了一种建立在数据源分析基础上对典型人脸识别算法进行后处理的方法 为了揭示现有典型识别算法的识别性能在无约束环境下的鲁棒性, 通过建立 Lambertian反射模型和 3D 人脸模型, 对 特征脸算法的识别性能随数据源的变化(人脸姿态和光照改变 )而变化的情况进行了分析评估针对/ 数据源灾难0 问题, 提出了一种基于隐马尔可夫模型 (HMM )的后处理解决方法, 该方法通过利用视频序列图像的连续性和对训练 人脸库的统计分析来提高判别分析方法对无约束环境的鲁棒性实验结果表明, 该方法可以有效地提高识别算法对 / 数据源灾难0的鲁棒性, 提高识别率 : 隐马尔可夫模型; 人脸识别; 后处理; 混淆矩阵; 先验模型; 置信度 : TP39 1. 4 1 : A App lication of post-processing based on HMM to video face recogn ition DA IY ,i X IAO Guo-qiang, SONG G ang (College of Computer and Information Science, Southwest University, Chongqing 400715, China) A bstract: In this paper, the rarely concerned prob lem of data source in face recogn ition w as investigated, and a novel post- processing HMM-based solution w as proposed. Data source problem w as firstly emp irically investigated through systematically evaluating the eigenfaces sensitivity to variations of pose and illum ination by Lambertian reflection model and 3D face m odel, which revealed that the changes of pose and illum ination abruptly degrade the eigenfaces system. Th is problem is e plicitly defined as data source disaster for h ighlighting its sign ificance. Ami ing at solving th is prob lem, com bin ing the recognition rate w ith the analysis of the data sources, wt om ethods were proposed to evaluate the overall performance of specific face recogn ition approach w ith its robustness against the low-quality data sources considered.