Novel Local Pattern Descriptors via Dynamic Linear Decision Function for Face Recognition

102

Recently, the research in face recognition has been focused on developing a face representation that is designed to generate invariant features for solving facial illumination and expression. Motivated by a simple but powerful local pattern descriptor, Local Binary Pattern (LBP), two novel local pattern descriptors are proposed to extend the LBP to vector-based and directional-based local pattern descriptors via dynamic linear decision function for face recognition. The first descriptor, namely, Local Vector Pattern (LVP), provides a novel vector representation and a coding scheme Comparative Space Transform (CST), which are used to generate more detailed discriminative local features than the other methods. The second proposed descriptor, namely, Local Directional Classifier Pattern (LDCP), computes eight edge response values from extra neighborhood pixels, and these values are used to select the upper and lower bound indices for generating robust complete binary codes. These methods are implemented and compared with existing LBP face recognition systems and other state-of-art local pattern descriptors on FERET, CAS-PEAL, CMU-PIE, Extend Yale B, and LFW databases. Experimental results demonstrate that the proposed methods outperform the other comparative methods with grayscale images and Gabor features as inputs.