Cascade shallow CNN structure for face verification and identification

Face recognition is a long-standing challenging topic in computer science, especially on insufficient datasets. The obstacle also lies in the balance of speed and accuracy. Recently, many algorithms claim that they have obtained great performance with high accuracy, but they are not enough for real-time application. In this work, a novel fast and accurate solution is proposed to deal with the face recognition problem on the small training set. Based on face alignment, we present two methods to extract features. One is a combination of several kinds of human designed feature descriptors applied on patches partitioned according to facial landmarks. The other one is a cascade classifier based on shallow convolutional neural networks. Both methods can represent the face as a set of feature vectors, which can be dealt with SVM or a boosting verification algorithm in this work. In the experiments, the proposed framework has achieved great performance for face recognition and verification with high speed and high accuracy, based on the public available datasets such as the Labeled Face in the Wild dataset and the AT&T database of faces.