Multiscale Block-Directional Local Binary Pattern (MB-DLBP) for Face Recognition in Illumination Variations

Majority of available descriptors do not possess the effective methodology under illumination variations. Their results are under par in such situations. So robust descriptor is needed in illumination changes. Therefore a novel local descriptor has been introduced in illumination variations. This novel local descriptor is MB-DLBP. In MB-DLBP, after mean computation from 6x6 block, the directional features are acquired by comparing the neighborhoods in clockwise direction. The features are derived globally and compaction is attained by FLDA. SVMs and NN are consumed for matching, the two effective techniques existing in literature. Datasets used are EYB and YB. Among three classifiers the best results are attained by the RBF classifier as compare to the other two, the POLY and ESA cosine. MB-DLBP outperforms several benchmark methods on both EYB and YB.