Robust local binary pattern for face recognition in different challenges

In Local Binary Pattern (LBP), the minute pixel difference introduces image noise as the virtue of which recognition rate decreases. Or in other words, LBP and most of the other descriptors possesses noisy thresholding function. To overcome this, the center pixel in 3 x 3 patch is replaced by the mean of whole patch. Then 2 descriptors are launched so-called Sign Binary Pattern (SBP) and Magnitude Binary Pattern (MABP). In SBP, the values induced from neighbors & mean (by taking difference) are absorbed by thresholding function for building the SBP size. In MABP, the absolute values induced from all pixels and mean (by taking difference) are availed by thresholding function for building MABP size. The SBP & MABP features are merged next for making discriminant feature called Robust LBP (RLBP). As SBP is characterized by mean comparison and MABP is characterized by absolute mean comparison, therefore their combination generate more robustness than various descriptors, which possesses noisy thresholding functions (proved experimentally in result section). Their combination also yields discriminativity against the other challenges such as light, pose, emotion and scale. SBP and MABP whole feature size are build by integrating features region wise. The methodology used for developing these descriptors is totally novel. As length of feature aligned on higher side so compression in feature length is gained by PCA and then SVMs and NN are used for matching. The RLBP ability is examined on ORL and GT by comparing with several methods. The proposed RLBP justifies its efficacy by conquering all of them. The numerous methods include those which are evaluated with proposed ones and they are LBP, HELBP and MBP. The rest compared methods are literature work.

Automated summary

In Local Binary Pattern (LBP), the small differences between pixels can cause image noise and decrease recognition rate. To overcome this, novel descriptors are introduced to improve robustness and discriminativity. These descriptors compare pixels in different ways and construct features based on thresholding functions. By combining these features, challenges such as lighting, pose, emotion, and scale can be addressed.