Learning compact discriminant local face descriptor with VLAD

Local Binary patterns (LBP) and its extensions typify the present face descriptors due to their intrinsic capability of featuring the neighborhood changes striding over every pixel. These descriptors are usually engineered in an obsolete handcrafted manner and thus sufficient prior knowledge and expertise are necessitated to assure the recognition performance. This paper outlines an improved face descriptor to the recently proposed learning-based discriminant face descriptor (DFD), coined compact discriminant local face descriptor (CDLFD). In general, the pixel discriminant matrices (PDMs) that store the LBP-like local intensity variations are pruned onto the discriminant pixel vectors (DPVs) with respect to the DFD learned feature filters and the optimal soft-sampling matrices. Different from DFD and other analogous state of the arts that cluster the extracted features into the bag-of-word representation, CDLFD encodes the DPVs as a set of vector of locally aggregated descriptors (VLADs). The global VLAD signature, i.e., the concatenation of all local VLADs, is appropriately normalized and PCA whitened to yield the globally compact representation. The CDLFD performance is scrutinized based on the standard FERET evaluation protocol and the AR dataset. The experimental results disclose that CDLFD outperforms the handcrafted LBP variants, DFD, and other face descriptors, in terms of rank-1 recognition rate (%).