Triangle and orthogonal local binary pattern for face recognition

LBP is known as one of the best performing local descriptor in texture representation. But there are various shortcomings observed in LBP and these are finite spatial patch and large feature size. These shortcomings also persist in numerous LBP variants. To remedy these shortcomings, the proposed work presents 2 LBP variants so-called Triangle LBP (TLBP) and Orthogonal LBP (OLBP), in pose and expression variations. TLBP features are extracted in horizontal and vertical directions by using 3 × 5 and 5 × 3 image patches, by rotating the triangle in 0 0 and 180 0 directions of both patch. OLBP features are extracted from orthogonal positions of the respective patch. The feature size derived from TLBP and OLBP descriptors are fused to manufacture the robust face descriptor called as Triangle And Orthogonal LBP (TAO-LBP). The compressed set of feature is accomplished by PCA. The finite spatial patch problem is eliminated by using two different patches, from which lower and higher scale features are extracted in the form of histograms, by using novel methodologies as introduced in TLBP and OLBP. The large feature size problem is scrutinized by the deployment of PCA and FLDA, which also selects the relevant and essential information for classification. The classification is procured by SVMs and NN. Experiments confirms the potential of TAO-LBP against LBP-like and non LBP-like based methods on ORL, GT, JAFFE, EYB and Faces94 datasets.