A mini-batch discriminative feature weighting algorithm for LBP - Based face recognition

RGB-D-T Based Face Recognition

“…The list of selected features is Local Binary Patterns (LBP) [15], Histograms of Oriented Gradients (HOG) [16] and HAAR-like features for face recognition [17]. The LBP is a very popular descriptor in the field of face recognition [15], [18]. The simplicity and high discriminative power of LBP in various computer vision tasks motivated the development of various extensions of the paradigm [19].…”

“…The final step of the pipeline is recognition. Three recognition strategies are tested here: the most simplistic one based on Nearest Neighbor Classifier (NNC), the Weighted NNC based approach [18], [19] and Linear SVM based principle with "One-vs-All" classification scheme. The algorithmic details are discussed later.…”

“…where K is the regioning factor (the LBP transformed image is divided into K × K regions), P is the number of sampling points on the radius R. The notation (P, R) is usually used for the description of LBP operator [15], [18].…”

“…In this instance the features are weighted according to their discriminative power prior the recognition. Following [18], [19] we have used an iterative feature weighting method called a Mini-Batch Discriminative Feature Weighting (MB-DFW) algorithm. The MB-DFW has a number of significant advantages: only two training samples per class are needed, which is important for biometric applications.…”

“…In the feature-level weighting each feature in the descriptive vector has a unique weight, while in the block-level weighting the features within the block have the same weight. According to the results in [18], [19] block-level weighting slightly outperforms the feature-level weighting concept.…”