Robust Multispectral Features For Face Liveness Detection

Face detection plays an important part in biometric recognition, human-computer interaction, and video surveillance systems. How to effectively improve the detection and recognition capabilities in complex environment is critical. The traditional human face detection is based on RGB images or videos, but the traditional method is easily interfered by man-like objects. Multispectral images with rich information can get more reliable results, which reduce the system false rates. This paper discusses the face detection technology based on narrow-band multispectral imaging system and this paper is built by the following aspects:1. The unique physiological characteristics of human skin is studied. The multispectral reflectance curves of human skin and skin-like object are measured and compared in400nm-1000nm wavelength by using a spectrometer.2. A narrow-band multispectral imaging system is built, which can collect the narrow-band multispectral images in different circumstance. The paper gathers the human face images (positive samples) and PVC mannequins, mask and face photos images (negative samples).3. In the situation of controllable brightness and specific shooting distance, the reflectance features of narrow-band images are extracted. The face liveness detection algorithm is proposed. After training, support vector machine in the proposed algorithm distinguish the face and face-like object effectively.4. In many practical scene, brightness and shooting distance are uncontrollable. The ratio-feature of narrow-band images is proposed to decrease the influence of the shooting distance. The difference-feature is proposed to eliminate the influence of the brightness. Then the paper proposes the difference-ratio feature and designs the face liveness detection algorithm with SVM. The experimental results show the difference-Ratio feature have the ability to eliminate the influence of brightness and shooting distance, effectively.The experiment results show the face detection based on multispectral imaging provides more advantage than traditional method, and has great potential in security region.