As an emerging human identification technique, face recognition has obtained great development, and has been successfully applied in many real world applications such as public surveillance and access control. However, these traditional face recognition techniques rarely consider the genuineness of target faces, which makes them vulnerable to fake face attacks. Considering that the success of attack may incur tremendous loss, it is urgent to design reliable and efficient liveness detection techniques for existing face recognition systems. This issue has now become a hot topic in face biometrics, and has attracted researchers all over the world. After reviewing the development in the literature, this thesis provides a systematic, thorough and deep analysis for the issue of face liveness detection. Basically speaking, this thesis focuses on the methods of multispectral imaging and multiple liveness clue fusion, and proposes efficient solutions respectively. To be specific, the main contributions of this thesis include: 1. Based on different reflectance properties, this thesis collects, analyzes and models the reflectance distribution of genuine and fake faces at various spectrums. The optimal spectrum selection criteria is discussed, and an efficient prototype system is given. 2. The problem of multispectral face detection is studied, which is further used as a rough but simple and efficient liveness detection strategy. As the insufficency of multispectral face images makes it difficult to train a multispectral face detector, this thesis proposes an efficient learning algorithm under the principle of transfer learning,which significantly boost the detection accuracy. Moreover, multispectral face detection is utilized as a simple face liveness detection method, and it achieves very good performance when anti-spoofing photo attacks. 3. For face recognition systems working under visible illumination, a face liveness detection method by fusing multiple liveness clues is proposed. Three kinds of liveness clues are designed for the IJCB'11 contest database, which are facial motion, face-background motion consistency and facial texture. The final decision is given based on the fusion of these three clues. The experimental results also verify the effectiveness of the proposed multiple liveness clues fusion method. 4. There usually lacks sufficient variation in fake face attacks for existing face liveness detection database. Aimed at this disadvantage, th...
修改评论