Improved LBP and Discriminative LBP: Two novel local descriptors for Face Recognition

Abstract

Most of LBP based descriptors develop their feature size by considering the uniform coordination among neighbors and center pixel. Additionally, most of them possesses noisy thresholding function. This limits the discriminativity of these descriptors to large extent. To eliminate all above defined conditions two novel descriptors are introduced called as Improved LBP (ILBP) and Discriminative LBP (DLBP). In ILBP, initially maximum value is attained from the 3x3 patch. Then product is taken between the maximum value and one of best possible values within range (.1-.9) to develop the threshold value. For this work it has been observed that.9 gives the best accuracy therefore.9 is used for obtaining the threshold value. The best range value will be chosen for obtaining the threshold value. Then all neighbors are compared against threshold for forming ILBP code. The ILBP histogram is achieved by computing the ILBP code for each pixel position. In DLBP, the histograms of ILBP and LBP are merged to form the more robust descriptor. For feature compression PCA is used and then classification was done by RBF technique, the SVMs method. Experiments conducted on 2 benchmark datasets i.e. ORL and GT confirms ability of both the descriptors against various methods. Among all it is DLBP which achieves best accuracy.