Research On Face Detection Method For Face Recognition Based On Face Recognition

Recently, face recognition technique has developed significantly, espe-cially when coping with pose, illumination and expression variations. There-fore, more and more applications adopt face recognition systems for identifica-tion or verification. However, most of recent face recognition systems can only recognize the identity of a face image, but cannot determine whether the source of face image, is a real person or not. Therefore, they are prone to various spoof-ing attacks, which may lead to severe outcome. To address this problem, face liveness detection (or face anti-spoofing) is proposed. This technique is aimed to determine whether the face image is captured from a real person or merely a photo. Based on efficient face liveness detection techniques, more secure and robust face recognition system can be achieved, which is indispensable in practical applications.After reviewing amounts of related papers, the author proposed several counter measures to the spoofing attacks based on the observations and analy-sis on the differences between real and fake faces. Besides, to make the live-ness detection module feasible in practical face recognition system, the author proposed a person-specific live face recognition framework. Specifically, the research works on face liveness detection in this paper mainly contain four sec-tions:1.Face liveness detection based on micro-texture:Generally, fabricating fake faces need more procedures, which may lead to the loss of details in the images. Moreover, the amounts of loss differ from region to region. In this paper, the author proposed a component-dependent descriptor for face liveness detection. Firstly, texture features extracted from a local face region is coded using specific codebook. Then, the codes are pooled to be a mid-level descriptor according to the output of Fisher Criterion. This method considers the differences of distinguishing ability among different face regions by imposing larger weights to the components with higher distinguish abilities and smaller weights to those with lower distinguish abilities. The extensive experimental results on three public databases demonstrate the effectiveness of the proposed method.2. Face liveness detection based on3-D face structure:This approach is based on the observation that a fake face printed on a photograph cannot mimic the structure of a real face. Therefore, the liveness of a face can be recognized by recovering its3-D structure. However, recovering a dense structure is significantly time-consuming. Alternatively, the author proposed to use the facial landmarks for recovering sparse3-D structure. After the sparse structure of a face image is recovered, it is fed into a binary classifier for liveness detection. This approach can effectively cope with the spoofing attacks from printed photos. The perfect experimental results on a database collected by ourselves show it efficiency.3. Face Liveness detection based on motion consistency:The micro-texture and3-D structure are not applicable for face liveness detection in some circumstances. For example, if the attacker display a face im-age using a high definition screen steadily, the above two approaches may be invalid. Beyond the micro-texture or3-D facial structure, the motion consistency between face region and background is another clue for face liveness detection. In general, there often exists a motion consistency be-tween the face region and surrounding background in a spoofing video. In contrast, this consistency is extremely slight or even void in a real access video. Based on this observation, this paper proposed a motion consis-tency based face liveness detection method. This method is very efficient when coping with some special situations.4. Person-specific face liveness detection:In the previous works, a generic anti-spoofing classifier is trained to detect spoofing attacks on all subjects. However, due to the individual differences among subjects, the generic classifier cannot generalize well to all subjects. Moreover, it is difficult to re-train the generic classifier every time new subjects are enrolled. In this paper, the author proposed a person-specific face anti-spoofing frame-work, which identifies input faces first, and then detects the spoofing at-tacks using the classifier trained for that specific subject. Another issue in practice is that it is unrealistic to obtain fake faces for all enrolled subjects. To train person-specific anti-spoofing classifiers for those subjects which have no fake samples, the author synthesized fake samples for them in the feature space alternatively. As shown in the experiments, by implement-ing person-specific face anti-spoofing combined with the feature synthesis algorithm, the anti-spoofing performance improves much.