隐马尔可夫后处理模型在视频人脸识别中的应用

Application of post-processing based on HMM to video face recognition

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,但缺乏一种对数据源(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"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"的鲁棒性,提高识别率。

In this paper,the rarely concerned problem of data source in face recognition was investigated,and a novel post-processing HMM-based solution was proposed.Data source problem was firstly empirically investigated through systematically evaluating the eigenfaces sensitivity to variations of pose and illumination by Lambertian reflection model and 3D face model,which revealed that the changes of pose and illumination abruptly degrade the eigenfaces system.This problem is explicitly defined as ＂data source disaster＂ for highlighting its significance.Aiming at solving this problem,combining the recognition rate with the analysis of the data sources,two methods were proposed to evaluate the overall performance of specific face recognition approach with its robustness against the low-quality data sources considered.Finally,a post-processing method was proposed to improve the robustness of the recognizer under unconstrained environment.The experimental results have impressively indicated the effectiveness of the proposed post-processing solution to tackle the ＂data source disaster＂ problem.