Abstract
This paper presents a face recognition method using a set of efficient local texture features, called multi-lag directional local correlations (MDLCs). They measure the intensity similarity between a local region and each of the counterparts which are multi-lag directional vectors distant from it, which is a sort of local correlation coefficient that is well normalized and bounded. Each of the MDLC images extracted from a facial image is then low pass filtered in the global 2D DCT (discrete cosine transform) domain, which reduces not only feature dimension but also noisy disturbance obstructing elaborate face recognition. The DCT coefficients retained from low pass filtering are all fused into a 1D feature vector for an input of a stabilized whitened cosine (SWC) distance classifier. The performance of the MDLC features is compared with those of Gabor wavelet, LBP (local binary pattern), gradient faces, the fusion of BDIP (block difference of inverse probabilities) and BVLCs (block variation of local correlation coefficients). Experimental results for six facial databases (DBs) with a single training image per person and with multiple training images show the MDLC features yield almost the best performance robust to variations of expression, lighting, and aging among the discussed features.