Abstract: Face recognition with single training sample per person, due to the lack of inner-class information to predict the complex variations, results in the recognition accuracy of commonly used algorithms declined sharply. And the best solutions can be broadly divided into two categories, one is to generate virtual training samples from original samples, the other is through learning a sparse variation dictionary to predict the variations. To solve this challenging problem, this paper proposes an algorithm based on automatic weights to the K nearest patches. Because the training dictionary which just has one training sample per class can not predict facial complex features change, this paper imports the sparse variation dictionary to represent the complex facial features changes. Then this paper divides the test sample into some sub-blocks, and picks out the K nearest patches of each sub-block to form a virtual testing set and gives automatic weights to those patches in the classification. Finally, this paper uses an improved voting mechanism to get classification results of the original test sample. Extensive experiments on representative face databases AR, CMU Multi-PIE and ORL demonstrate that the proposed algorithm is much more effective than state-of-the-art algorithms in dealing with face recognition with single training sample per person.

Key words:  single training sample, face recognition, sparse variation dictionary, K nearest patches, automatic weight, voting mechanism

摘要： 在人脸识别问题中，当每类训练样本有且仅有一个时，由于类内缺乏足够的特征变化信息来预测人脸复杂的特征变化，从而导致常用分类算法的识别准确率急剧下降。目前最好的解决方法大致可分为两类:一是生成虚拟的训练样本以扩大训练集；二是学习稀疏变化字典以表示复杂特征变化。针对此问题，在引入稀疏变化字典来表示人脸复杂特征变化的基础上，提出一种基于K邻域分块自动加权的单样本识别算法。通过对测试样本进行分块，然后对每一个子分块求K邻域分块，以组成虚拟的同类别测试样本集；同时提出了一种自动加权策略，对这些分块在分类中的比重进行加权，最后通过一种改进的投票机制确定分类结果。通过与已有的单样本识别算法进行比较，并在公共人脸数据库AR、CMU Multi-PIE和ORL上进行实验，结果表明该方法有助于提高单样本识别问题的分类准确率。

关键词: 单训练样本, 人脸识别, 稀疏变化字典, K邻域分块, 自动加权, 投票机制