Abstract: A new algorithm based on neighborhood repulsed metric learning sparse discriminant for single sample is proposed to solve the problem of kinship verification with facial image. Firstly, a new distance metric is learnt under which facial images with kinship relations are projected as close as possible and those without kinship relations in the neighborhoods are pulled as far as possible. The purpose is to use the similarity of existing data samples to learn a distance metric which can better describe the similarity of samples. Then, the sparse representation method is adopted to establish a dictionary which can be used to linearly represent the children images under the new distance metric space, and the sparse coefficient is applied to measure the similarity of different samples. In addition, to deal with the inconspicuous similarity of kinship samples, a novel algorithm based on sub-modular sparse discriminant is proposed. Finally, the multiple sparse coefficients are used to decide whether there is a kinship relation between the two input samples. The experiment results on the KinFaceW-I database and KinFaceW-II dataset demonstrate the efficiency of the proposed method.

Key words: kinship verification, metric learning, sparse representation, similarity metric

摘要： 针对如何利用人脸图像进行亲属关系认证问题，提出邻域排斥测度学习稀疏判决的单样本亲属关系认证算法。学习能使具有亲属关系样本距离变小，而非亲属关系样本距离变大的变换矩阵，目的是利用已有数据样本间相似程度的先验知识学习最佳相似性度量，使之能更好地刻画亲属样本间的相似关系。在新的测度空间下采用稀疏表示方法用父母样本集建立过完备字典来线性表示子女图像，并以稀疏系数大小衡量样本间相似程度。针对亲属样本间相似性不明显问题提出子模块综合稀疏认证方法，通过多重稀疏系数综合判别两输入样本的亲属关系。在KinFaceW-I和KinFaceW-II两个亲属图像库上的实验结果表明，采用测度学习空间下稀疏系数判决的方法相比已有亲属关系人脸认证方法具有更好的性能。

关键词: 亲属关系认证, 测度学习, 稀疏表示, 相似性度量