Deep Frequency Quantization for Face Image Recognition and Liveness Detection

Image representation plays a critical role in face image recognition and liveness detection. In this paper, we propose a novel deep learning-based image feature extraction method, termed deep frequency quantization (DFQ), for face image recognition and liveness detection. First, the phase and magnitude features of the image are extracted by applying Fourier transform and short-time Fourier transform, respectively. Second, stacked auto-encoders (SAEs) are employed to learn high-level representations of the Fourier phase and magnitude features. Third, separate codebooks for the deep Fourier phase and magnitude feature vectors are constructed using Lloyd’s algorithm. Each feature vector is subsequently matched to the nearest codeword within its corresponding codebook, and the index of this codeword is defined as the DFQ pattern. Finally, histograms of the phase and magnitude DFQ patterns are generated and concatenated to form the feature representation of the entire image. Extensive experimental results on six benchmark databases demonstrate that the proposed DFQ method consistently outperforms state-of-the-art approaches for face recognition and liveness detection.