Face anti-spoofing detection based on multi-scale image quality assessment

Abstract

This article introduces an effective face PAD algorithm based on multiscale perceptual image quality assessment features. Unique hand-crafted texture features extracted from face images are exploited for spoofing detection. The proposed features are classified into three major models: generalized Gaussian density-based, asymmetric generalized Gaussian density-based, and top gradient similarity deviation features. In light of the essential attributes of these models, a total of 21 multiscale features are acquired for classification, which is performed through a support vector machine (SVM). Extensive experiments on five benchmark databases, CASIA, Replay-Attack, UVAD, OULU-NPU, and SiW along with our new dataset demonstrated the effectiveness of the proposed framework. Experimental results indicated that our face PAD algorithm produced satisfactory detection accuracy on the tested datasets based on both intra-dataset and cross-dataset protocols. While outperforming a number of traditional face PAD methods, the proposed scheme achieved comparable results with many state-of-the-art deep learning-based networks. The introduction of the image quality assessment features with multiscale analysis into face PAD is promising for detection accuracy improvement.