在生物特征识别领域，人脸识别研究一直是一个关注的热点。目前最好的自动人脸识别系统在注册和认证环境条件比较一致、用户比较配合的情况下已经能够达到令人满意的效果。但是，在采集环境不可控、用户不配合以及人脸库规模过大的情况下，人脸识别算法和系统的性能就会急剧下降。人脸识别研究还面临许多待研究的课题，其中包括如何设计基于统计学习的分类器和开发消除光照影响的方法。我们经过总结和借鉴已有的工作，在上述两个问题上发展出了三个新颖的算法： 1. 提出了杠杆学习机（Lever Training Machine）。该算法受到物理学中的杠杆原理启发，将力矩平衡的概念引入到高维空间中，并利用该原理优化地训练线性分类器。从理论上的推导证明，杠杆学习机的优化目标直接就是最小化分类器的误识率。而且在目标样本是凸分布的情况下，可以利用杠杆学习机训练一系列优化的分界面来将目标样本分布与外界隔离开来。我们在二维和三维欧氏空间上直观地验证了杠杆学习机的有效性，并进一步将该算法成功应用到了人脸检测上面。在人脸检测应用中，还发现杠杆学习机具有学习面部样本全局特征的能力。该算法的成果已发表在ICNC2005论文集中。 2. 提出了相对差分空间变换（Relative Difference Space）。通过引入参考点的概念，该算法避免了差分空间的病态变换问题，并且具有可逆性。相对差分空间变换维持了类间的可分性，并可以在一定程度上缓解训练过程中的小样本问题。在此基础上，我们基于该算法和支持向量机构建出了一种新颖的多类分类方法，并成功应用于处理人脸识别的光照问题。在Yale Face Database B库上的测试结果证明我们的算法能有效地消除光照对人脸识别结果的影响，其性能接近甚至超过了现有的专门处理光照问题的算法，相关成果已经发表在ICPR2006论文集中。 3. 提出了形态学商图像算法（Morphological Quotient Image）。我们根据面部光照特点，采用数学形态学算子和商图像技术对各种光照条件下的人脸图像进行归一化处理。与传统的技术相比，该方法不需要训练数据集以及假定光源位置，并且每人只需一幅注册图像。进一步，我们通过引入特征尺度测量参数和动态模板技术，提供了一套复杂光照条件下的动态光照补偿办法，也即动态形态学商图像算法。Yale Face database B上的实验结果有力地证明了我们算法在处理复杂光照条件方面的性能，与现有的算法相比，形态学商图像算法在时间和空间复杂度上同样具有优势。目前，该算法的成果已被软件学报接收。
英文摘要
Face recognition is an important subject of Biometrics. At present, the best Automation Face Recognition System performs well with strictly controlled environment and good customer cooperation, but the efficiency of face recognition system and algorithm will drop down when the work condition changes variously, customer has no patience, or the face database becomes very large. Therefore there are still many hard subjects to be handled. Statistical learning based classification and illumination compensation are two open problems in face recognition. With the experiential lessons of the existing works on the two subjects, we proposed three novel algorithms as below. 1. Lever Training Machine. A new learning algorithm, Lever Training Machine (LTM), is proposed to construct a novel linear binary classifier. The idea of LTM is inspired by Lever Principle in physics, and LTM introduces the principle to hyper dimensional space for optimizing the decision surface. In theory, it is proved that LTM can minimize the false positive rate. If the target distribution is convex, a set of such decision surfaces can be efficiently trained by LTM for exact discrimination, which is conformed by experiments in 2D and 3D Euclid space. Finally LTM is applied for face detection successfully and it is found that LTM can effectively learn the holistic features of face images. This study has been published in proceedings of ICNC, 2005. 2. Relative Difference Space. By employing reference point, Relative Difference space (RDS) overcomes the ill-transformation problem of Difference Space (DS) and it becomes reversible. RDS keeps well the structural information of original patterns and can alleviate small sample size problem. Incorporating RDS with Support Vector Machine (SVM) together, a novel method is proposed for multi-class recognition. This method is applied to face recognition under various illumination conditions. The recognition result on Yale Face database B demonstrates its robust performance of handling illumination problem as effectively as those special algorithms for illumination normalization. This work has been published in proceeding of ICPR, 2006. 3. Morphological Quotient Image. An effective illumination normalization method is proposed based on mathematical morphology and quotient image techniques by analyzing the face illumination. Compared with traditional approaches, this method does not need any training data and any assumption on the light conditions, moreover, the enrollment requires only one image for each subject. With dynamical lighting estimation technique, the proposed algorithm is upgraded to strengthen illumination compensation and feature enhancement. The proposed methods are evaluated on Yale Face database B and receive a very competitive recognition rate with low computational cost. This research has been accepted by Journal of Software recently.
修改评论