关键词: 局部二值模式, 局部序数模式, [K]均值, 白化主成分分析

Abstract: Local Binary Pattern（LBP） encodes images by comparing 8 neighborhoods pixels to the central point which ignore the magnitude of neighborhoods, Local Intensity Order Pattern（LIOP） encodes neighbor pixels’ local ordinal information which works only in small numbers of neighbor pixels, aiming at these disadvantages, a new feature extraction algorithm with large number of neighbor pixels’ local ordinal information is proposed. The algorithm firstly gets the feature vectors which contain neighborhoods ordinal information, then uses the unsupervised clustering method（[K]-means） to reduce pattern numbers of feature vectors. This clustering process can be done offline and run efficiently. Finally, cascaded region histogram features of images are used as the final face representation. Experimental results show that ELQP descriptor achieves the best robustness and recognition rate among these compared descriptors. Furthermore, Whitened Principal Component Analysis（WPCA） is applied to get less feature dimensions and better recognition rate.

Key words: Local Binary Pattern（LBP）, Local Intensity Order Pattern（LIOP）, [K]-means, Whitened Principal Component Analysis（WPCA）