对单训练样本的人脸识别问题的研究

Some Researches for Face Recognition with One Training Image per Person

现在许多人脸识别算法都是在假定每个人提供了多幅训练样本的情况下展开的,对每人只有一幅训练图像的识别问题研究得很少,而实际中往往每人只提供了一幅图像。本文对这一问题进行了研究,给出了一些生成虚拟训练样本的方法;提出了基于类间散度最大的二维主分量分析方法,在 ORL 库上用单训练样本取得了75.28%的识别结果。

Nowadays many algorithms for face recognition are under the postulate that each person has many training images. There are few study with the one training sample per person. While each person may only provide one registered photo in most cases. We solve this problem by add virtual images generating from the given training image, and study the differences of the recognition rates between PCA, Fisherface, （PC）^2A and Two Dimension PCA（2DPCA）. In this paper, a new 2DPCA which is based on the Maximum Margin Criterion is proposed. The average recognition rate on ORL face-databases achieves 75.28% only using one training image per person.