科研之友
微信
新浪微博
Facebook
分享链接
摘要
The spatially enhanced local binary pattern (LBP) histogram (eLBPH) methodology has attained an established position in the field of face recognition (FR) and derived many face analysis approaches. Their implementations follow a similar procedure: first divide a full facial image into some regions (subimages) and individually extract LBP histogram for each region, then concatenate all these regional histograms into a single (global) histogram for final recognition. It has been reported that eLBPH is more effective than the naive holistic LBP histogram (hLBPH), while the adoption of holistic LBP image (hLBPI) in FR is relatively few. So, this paper aims to systematically empirically address these issues: (1) Why the simple hLBPH is hardly adopted in FR? (2) Why eLBPH is more effective than hLBPH for FR? (3) hLBPI enjoys what kind of properties for FR. Concretely, we (1) compare the hLBPHs for large-variational facial images with those for standard texture images, and suggest that the LBP histogram feature generally needs certain preprocessing or post-processing for good FR performances; (2) illuminate the reason that eLBPH is more effective than hLBPH for FR, i.e., the enhanced histogram tends to be uniform (more stable than the holistic histogram) and relatively preserve spatial relations of faces, and show the sensitivity of eLBPH to the division region parameter; (3) we study the properties of hLBPI for FR, i.e., hLBPI faithfully preserves the both spatial structure and intrinsic appearance details of a facial image, inherits the attractive properties of the LBP operator and does not require the calculation of histogram for FR; (4) comprehensively evaluate and compare hLBPI, hLBPH, eLBPH and some subspace algorithms on the benchmark face datasets (FERET, Extended YaleB, CMU PIE, AR); (5) conclude that hLBPI, hLBPH and eLBPH respectively are suitable for face representation under what conditions, and expect providing practitioners with guidance in selecting appropriate approaches for real tasks.
