Binarized features with discriminant manifold filters for robust single-sample face recognition

Many popular face recognition methods perform excellently when multiple samples per person are available for the training step of feature extraction. However, in many practical applications, there might be only a single sample recorded in systems for each person and most of traditional methods perform not very well because they usually need much more samples for discriminant learning. To address this problem, in this paper we propose a novel method for constructing local histogram based facial image descriptors, inspired by discriminant manifold learning and binary encoding. In manifold learning procedure, we divide each image into several nonoverlapping patches and try to find a matrix to project these patches onto an optimal subspace to maximize manifold margins of different persons. Then each column of the matrix is reshaped to an image filter to process facial images and the responses corresponding to these filters are binarized via thresholding. Thus for each pixel we can obtain a binary code whose length is just the number of filters. Finally, we compute region-wise histograms of pixels' binary codes and concatenate them to form the representation of a facial image for recognition. We conduct some experiments on Extended Yale-B, AR, FERET and LFW databases to demonstrate the effectiveness of our proposed method in face recognition.