Face anti-spoofing via deep local binary pattern

Abstract

In recent years, convolutional neural network (CNN) has achieved satisfactory performance in computer vision and pattern recognition. When we visualize the convolutional responses, we can conclude that the convolutional responses include some diacritically structural information. But for the high dimensionality of them, it is not feasible to directly use the responses to detect fake faces. Moreover, the small size of existing face anti-spoofing databases leads to the difficulty of training a new CNN model. Compared with deep learning, the traditional handcrafted features, such as local binary pattern (LBP), have been successfully used in face anti-spoofing and achieved good detection results. So in our work, we extracted the handcrafted features from the convolutional responses of the fine-tuned CNN model. More specifically, the CNN is first fine-tuned based on a pre-trained VGG-face model. Then, the LBP features are calculated from the convolutional responses and concatenated into one feature vector. After that, the final vectors are fed into a support vector machine (SVM) classifier to detect the fake faces. Validated on two public available databases, Replay-Attack and CASIA-FA, our proposed detection method can obtain promising results compared to the state-of-the-art methods.