结合加权子空间和相似度度量学习的人脸验证方法研究

在无约束条件下,人脸表情、姿态、光照以及背景等复杂因素可能导致人脸图像的类内变化大于类间变化.针对如何降低较大的类内变化对人脸验证研究的影响,本文结合加权子空间,提出了一种带先验相似性和先验距离约束的相似度度量学习方法.首先,利用类内人脸对样本,学习带权重的类内协方差矩阵,通过加权子空间的投影,从人脸图像中获得鲁棒性的人脸特征表达;其次,利用样本对的相似性与差异性,建立了带先验相似性和先验距离约束的相似度度量学习模型,优化后的度量矩阵可以有效提高特征向量的类内鲁棒性和类间判别性;最后,利用优化的度量矩阵计算人脸对的相似度.在LFW(Labeled Faces in the Wild)数据集的实验验证了所提模型的有效性,与其它同类相似度度量学习方法相比,优化的度量矩阵更能准确地评估人脸间的相似性,并在受限训练集上取得了91.2%的识别率.

Face Verification Based on Weighted Subspace and Similarity Metric Learning

Under the unconstrained conditions, intra-personal variation is much larger than the inter-personal variation in face images due to the affecting factors such as expression, posture, illumination and background etc. To reduce the influence of larger intra-personal on face verification, we proposed a similarity metric learning method with priori similarity and priori distance constraint by combining weighted subspace.First, the weighted intra-personal covariance matrix is learned by employing intra-personal face samples. By projecting into the intra-subspace, robust face feature representations can be obtained from face images.Second, we set up the similarity metric learning model with priori similarity and priori distance constraint, which effectively employs the similarity and discrimination information of samples that are in pairs, and the learned metric matrix can improve the robustness to intra-personal and discrimination to inter-personal.Finally, the updated metric matrix is used to compute the similarity scores of face-pairs. The experiments have been conducted on the Labeled Faces in the Wild (LFW) dataset, which shows the effectiveness of our proposed model. Compared with other metric learning methods, our learned metric matrix has higher accuracy rate for evaluating the face-pair similarity, and achieves a verification rate of 91.2% on the restricted setting.