Local features by intensity increasing tendency for face recognition

摘要

Local binary patterns (LBP) histogram has gained popularity as face descriptor. However, LBP is limited by its high dimensionality (256). The uniform local binary patterns (ULBP) reduces dimension to 59 at the price of sacrificed performance. In practice, both LBP and ULBP use neighbourhood with no more than eight pixels. This paper proposes a concept named increasing intensity vector (IIV), which specifies local intensity increasing tendency. Based on IIV, two novel local features are proposed: 1 local binary IIV (LBIIV), which extracts IIV from a binarised neighbourhood 2 local ternary IIV (LTIIV), where the extracted IIV is expressed in ternary mode. Compared to LBP and ULBP, LBIIV has ever fewer patterns (37) without degrading performance, and greatly improves efficiency. Meanwhile, using larger neighbourhood becomes practical. LTIIV then studies an insight into ternary concept of local pattern. Face recognition experiments show that LBIIV and LTIIV outperform LBP and ULBP in accuracy and efficiency.