Face recognition with single sample per person using HOG–LDB and SVDL

Abstract

The recognition rate of some face recognition methods that require a certain number of samples will be significantly reduced in case only one sample is available for training. Aim at this situation, a new feature extraction method, HOG–LDB (histogram of oriented gradients–local difference binary), is proposed. Then, we combined this method with SVDL (sparse variation dictionary learning) to recognize the probe images with different facial variations (e.g., illuminations, poses, expressions and disguises). The descriptor of HOG–LDB can extract the edge features and local pattern features of the image. After the feature extraction, SVDL is employed in the generic training set and the generic variation dictionary is obtained. Then, the dictionary is used for predicting the subjects of the probe images with different facial variations. Finally, experimental results on the AR dataset, the Yale dataset, the Extended Yale B dataset and the CMU-PIE dataset proved the validity of the proposed method.

http://bit.ly/2SHPctX