Feature extraction using block-based local binary pattern for face recognition

It is widely assumed that texture is generally characterized locally by two complementary aspects, a pattern and its strength. Based on this assumption and using Local Binary Pattern (LBP) operator as texture descriptor, this work aims to implement an automatic weighting of the local blocks or regions characterizing a given face image. The work reports an improved version of the margin-based iterative search Simba algorithm to feature extraction for face recognition. The main contribution is twofold: (i) we extend the margin-based iterative search algorithm (Simba) to the Chi-square distance that computes dissimilarities between histograms. (ii) since we are interested in studying the relevance of individual blocks or local regions characterizing a given face image, we also extended the Simba algorithm so that one can compute the weights of each attribute as well as of subsets of attributes or blocks. The resulting weight vector has been used initially for an automatic selection of attributes and/or blocks for face recognition with supervised learning based on k-nearest neighbors classifier. Besides, in order to improve the performance of the face recognition task we also made use of the Simba weight vector to weight the distance measures adopted by the k-NN classifier. The experimental results clearly show that the selection based on the automatic weighting outperforms the classification based in all the features. Furthermore, selecting blocks is more effective than selecting attributes, and Chi-square distance performs appreciably better than Euclidean one.