In this paper, we propose a face anti-spoofing strategy by using DWT (Discrete Wavelet Transform), LBP (Local Binary Pattern) and DCT (Discrete Cosine Transform) with a SVM classifier to evaluate whether a video is valid. Firstly, the DWT features are produced by decomposing some selected frames into different frequency components at the $8\ast 8$ multi-resolution blocks. Secondly, the DWT-LBP features are generated to represent spatial information of the blocks by connecting LBP histograms of the DWT blocks in each frame horizontally. Then, the DWT-LBP-DCT features with the temporal information of a video file are achieved by performing DCT operation on the DWT-LBP features of those selected frames vertically. As a result, these exploited DWT-LBP-DCT features have the capacity to represent the frequency-spatial–temporal information of a video. Finally, the SVM classifier with RBF kernel is trained for face anti-spoofing. Compared with previous excellent works, experimental results on two benchmark databases (REPLAY-ATTACK and CASIA-FASD) have demonstrated the proposed approach has better detection performance.

在本文中，我们提出了一种使用DWT（离散小波变换），LBP（局部二进制模式）和DCT（离散余弦变换）以及SVM分类器的人脸反欺骗策略，以评估视频是否有效。首先，DWT特征是通过将一些选定的帧分解为不同频率分量而产生的。$8\ast 8$多分辨率块。其次，通过水平连接每个帧中DWT块的LBP直方图，生成DWT-LBP特征以表示块的空间信息。然后，通过在垂直方向上对那些选定帧的DWT-LBP特征执行DCT操作，来获得具有视频文件的时间信息的DWT-LBP-DCT特征。结果，这些利用DWT-LBP-DCT的功能可以表示视频的时空信息。最后，对带有RBF内核的SVM分类器进行人脸反欺骗训练。与以前的优秀作品相比，在两个基准数据库（REPLAY-ATTACK和CASIA-FASD）上的实验结果表明，该方法具有更好的检测性能。