基于领域适应流形度量学习的无约束人脸识别方法研究

Unconstrained face images show multiple variances on illumination, expression, pose, blurring, alignment, etc. and are characterized difference between training set and test set, both of which will result in performance degradation when applying traditional metric learning methods to classify them. In order to solve the problem mentioned above, it will be expected to carry out the following works: 1) A label similarity based spectrum projection model will be constructed to explore a new domain with robustness to unconstrained face image variances, then a least locally linear reconstruction metric based on structure sparsity is put forward to identify them; 2) A domain model named Grassmann product manifolds is recommended to represent the multiple variances of unconstrained face images, based on which a log-Euclidean distance metric, which can measure geodesic distance on Grassmann product manifolds, is presented to predict the class of unlabelled data set; 3) A maximum mean discrepancy(MMD) in reproducing kernel Hilbert space is introduced to quantify the difference between training set and test set, moreover, a multiply manifold distance matrix is modeled using affine Hull representation, at last a manifold metric is reasoned to learn an discriminant adaptive projection domain, where k nearest neighbors can be adopted for unconstrained face classification. On one hand, the project will contribute to develop new manifold metric learning methods based on domain adaptation; on the other hand, the proposed methods will help to improve the efficiency of unconstrained face recognition.

无约束人脸图像具有光照、面部表情、姿态、模糊和对准等多种变化以及其中训练数据与测试数据存在差异，都将导致传统度量学习方法对无约束人脸图像的识别效率下降。本项目为解决以上问题，拟开展如下研究：设计一种基于类别相似度度量的谱映射模型，寻找对人脸图像各种变化鲁棒的变换领域，提出基于结构稀疏的局部最小线性重构标准进行分类；建立表征无约束人脸图像多种变化的Grassmann积流形领域模型，提出Grassmann积流形对数欧氏距离，作为积流形测地距离度量，预测人脸图像类别；利用再生核Hilbert空间最大均值差模型表示训练数据与测试数据分布差异，建立基于仿射Hull表示的多流形距离模型，研究适应判别领域学习的流形度量，寻找最佳分类投影领域，应用K近邻标准分类。本项目的研究一方面将促进基于领域适应的流形度量学习方法的发展，另一方面还能提高无约束人脸识别的效率。

无约束人脸图像具有光照、面部表情、姿态、模糊和对准等多种变化以及其中训练数据与测试数据存在差异，都将导致传统度量学习方法对无约束人脸图像的识别效率下降。本项目为解决以上问题，开展了如下研究：设计一种基于类别相似度度量的谱映射模型，寻找对人脸图像各种变化鲁棒的变换领域，提出基于结构稀疏的局部最小线性重构标准进行分类，在此基础上，提出基于特征空间到特征空间距离学习方法实现无约束人脸识别，提高了模型对姿态、表情、光照变化的鲁棒性能；建立表征无约束人脸图像多种变化的Grassmann积流形领域模型，采用Grassmann积流形的对数欧氏距离作为流形测地距离度量，作为领域适应学习的度量标准，实现了积流形的几何感知学习，预测人脸图像类别；利用再生核Hilbert空间最大均值差模型表示训练数据与测试数据分布差异作为领域适应学习的优化目标，建立基于仿射Hull表示的多流形距离模型表示多流形距离，研究适应判别领域学习的流形度量，寻找最佳分类投影领域，应用K近邻标准分类。另外本项目还研究了各种基于局部几何感知的领域适应度量标准，作为损失函数，应用于卷积神经网络的领域适应学习，提高无约束人脸识别的效果。本项目的研究一方面将促进基于领域适应的流形度量学习方法的发展，另一方面还能提高无约束人脸识别的效率及在安防监控等领域的应用。