An Anti-Photo Spoof Method in Face Recognition Based on the Analysis of Fourier Spectra with Sparse Logistic Regression

Among others, spoofing with photos is one of the most common manner to intrude a face recognition system. In this paper, we presents a novel method to deal with this problem, based on the observation that the difference between a photo and a real face usually leads to different distribution behavior in the frequency domain. In particular, we propose to first use a DoG (Difference of Gaussian) filter on the given image to preserve rich information for the subsequent stages while suppressing less-discriminative energy as much as possible. A two- dimensional discrete Fourier transformation is then applied on the filtered image, which produces an input to a sparse logistic regression model to give the final determinant. Furthermore, we adopt an early stoping strategy to prevent the logistic model from overfiting when our training set has class imbalance problem. We also investigate the influence of degree of sparsity on the performance of the system. Extensive experiments on a large scale testing set verify the feasibility and effectiveness of the proposed method.