Adaptive Semi-supervised Marginal Fisher Analysis
自适应半监督边界费舍尔分析

Graph based semi supervised methods have successfully used in face recognition. These algorithms not only consider the label information, but also utilize a consistency assumption. Conventional algorithms assumed that the consistency constraint is defined on the original feature space. However, the original feature space is not the best for defining consistency. We proposed adaptive semi supervised marginal fisher analysis(ASMFA) by which the consistency constraint is defined in the original feature space and the expected low-dimensional feature space. Experimental results on the CMU PIE and YALE-B databases demonstrate that ASMFA brings signification improvement in face recognition accuracy.