Image Quality Assessment Based Outlier Detection for Face Anti-Spoofing

Planar spoofing is a well researched problem, wherein a high quality planar photograph can be replayed in front of a still camera as a substitute for another individual's face. Most modern day face recognition systems can be fooled by this process, as the perceptual information contained in a photo-of-a-photo, is virtually the same as that of a natural photograph of an individual. Current solutions attempt to detect this form of planar-spoofing through an extrinsic training process wherein both planar samples as well as regular photos are included as separate training sets. To avoid this form of explicit discriminant model-learning, we propose a single class training procedure for establishing and quantifying the quality of natural photographs taken under different lighting conditions, in terms of their CONTRAST PROFILE. Once this distribution is learnt, a suitable threshold is set based on the mean and standard deviation to pick up outliers. In this paper, we show that with just single poses of subjects, it is possible to achieve a low Equal Error Rate (EER) of 21.56% on the CASIA dataset and a rate of 8.57% upon cross-validation with a trimmed and shortened version of the MSU dataset.