An Improved Face Liveness Detection Algorithm Based on Deep Convolution Neural Network

Abstract

Face recognition system has been frequently used in people's daily life. Therefore, face anti-spoofing technology on video and photo has attracted more and more attention. Based on the traditional VGG-11 model, we proposes an improved deep convolution neural network which can accurately detect the face liveness of single face image. Firstly, the training data set is enhanced by some methods such as random rotation, random brightness and saturation adjustment, which can improve the generalization ability of the network. Secondly, batch normalization and random deactivation are added into the traditional VGG-11 network, which can improve feature extraction and decision-making classification of real and false face images. Finally, we use the exponential attenuation learning rate during the process of network training which can avoid the optimization parameters wandering around the local optimization values. Experimental results against the state-of-the-art methods on NUAA and CASIA face liveness databases show that, the proposed method can achieve higher accuracy and lower recognition error rate on face liveness detection.