Weighted Multi-scale Local Binary Pattern Histograms for Face Recognition

This paper proposes a novel face recognition methodology, which is based on the combination of the texture operator, namely Multi-scale Local Binary Pattern (MSLBP), face image filtering and discriminative feature weighting algorithms. Presented MSLBP principle enhances the discriminative power of the original Local Binary Pattern operator. The combination of the MSLBP and low-pass filtering improves the stability of the feature vector for different scales of the input face. Proposed mini-batch discriminative feature weighting methodologies are applied to the feature space in the feature and block levels in order to enhance the components more relevant to the recognition process. The observed PCA based data compression algorithm significantly reduced the dimensionality of the feature vector with a minimal loss in the precision of the recognition process. The identification precision exceeded the threshold of 99% for the frontal subsets of a color FERET database.