Semi-Supervised Palmprint Recognition Based on Similarity Projection Analysis

Similarity is one of the most widely used measures in the field of pattern recognition like Euclidean and Mahalanobis distances. Semi-supervised learning is an effective technique for feature extraction, which can make full use of the unlabeled samples for training. In this paper, we incorporate similarity into semi-supervised learning and propose a novel feature extraction approach, named semi-supervised similarity projection analysis (SSP), for palmprint recognition. SSP projects original samples from a high-dimensional space to a low-dimensional subspace in a semi-supervised manner. It can preserve the similarity between intra-class samples and the dissimilarity between inter-class samples, and simultaneously maintain the global dissimilarity among both labeled and unlabeled samples. Experimental results on the HK PolyU palmprint image database demonstrate that the proposed approach outperforms several representative unsupervised, supervised and semi-supervised subspace learning methods.