对表情鲁棒的面部轮廓线3维人脸识别

Expression-robustness 3D face recognition based on facial profiles

目的表情变化是3维人脸识别面临的主要问题。为克服表情影响,提出了一种基于面部轮廓线对表情鲁棒的3维人脸识别方法。方法首先,对人脸进行预处理,包括人脸区域切割、平滑处理和姿态归一化,将所有的人脸置于姿态坐标系下;然后,从3维人脸模型的半刚性区域提取人脸多条垂直方向的轮廓线来表征人脸面部曲面;最后,利用弹性曲线匹配算法计算不同3维人脸模型间对应的轮廓线在预形状空间(preshape space)中的测地距离,将其作为相似性度量,并且对所有轮廓线的相似度向量加权融合,得到总相似度用于分类。结果在FRGC v2.0数据库上进行识别实验,获得97.1%的Rank-1识别率。结论基于面部轮廓线的3维人脸识别方法,通过从人脸的半刚性区域提取多条面部轮廓线来表征人脸,在一定程度上削弱了表情的影响,同时还提高了人脸匹配速度。实验结果表明,该方法具有较强的识别性能,并且对表情变化具有较好的鲁棒性。

Objective Given the non-intrusive nature and broad surveillance application of face recognition, this technology has drawn considerable attention in the fields of pattern recognition and computer vision. However, expression variation is one of the main challenges in 3D face recognition because the geometric shape of a face changes drastically under expression variation. For instance, an open mouth can significantly change the topology of the facial surface, which can degrade the performance of a 3 D face recognition system. To handle facial expressions, a novel 3 D face recognition meth- od based on facial profiles is proposed. Method First, the pose of a cropped face is automatically corrected on the basis of principal component analysis, and all facial scans are transformed into a uniform pose coordinate system. A set of ver- tical facial profiles in the upper half face region is then extracted to represent a 3D facial scan. Hence, the shapes of two facial scans can be matched by fitting the shapes of the corresponding facial profiles. Open curve analysis algorithm is applied to calculate the geodesic distance between a pair of facial profiles extracted from different facial scans. The geodesic distance is used as a similarity measure. Finally, two facial scans can be matched by using the weighted sum of all levels of the corresponding geodesic distance. Result One of the large stavailable public domain 2D and 3D human face data- sets is the Face Recognition Grand Challenge （ FRGC ） v2.0, which has been widely used in the literature. Two experiments are conducted using the FRGC v2.0 dataset： recognition and expression robustness experiments. In the recognition experiment, the earliest neutral 3 D facial scan of every individual is selected to create a gallery of 466 facial scans, and the rest are used as probes. We test three dataset partition methods that are commonly used in existing 3D face recognition systems, which also use FRGC v2.0 as the testing dataset （ i. e. , non-neutral vs. neutral, all vs. neutral, and neutral vs. neutral）. The Rank-1 recognition rates of our proposed approach in the cases of non-neutral vs. neutral, all vs. neutral, and neutral vs. neutral are 95.2% , 97. 1% , and 98% , respectively. In the expression-robustness experiment, we consider the gallery in the recognition experiment, and 816 facial scans with an open mouth from the FRGC v2.0 dataset are used as the testing set for face recognition. When the facial profiles are extracted from all the facial regions as features, the Rank-1 recognition rate is 82.8% , whereas that of our proposed method is 93.5%. Conclusion Achieving high accuracy in the presence of expression variation is one of the most challenging aspects of 3 D face recognition. To address this problem, a 3D face recognition method based on facial profiles is proposed. A set of vertical facial profiles are then extracted to represent facial surface. Given that these facial profiles are extracted from the semi-rigid region of a face, our proposed approach weakens the adverse effects caused by facial expression, especially large facial expression deformation, and consequently improves the efficiency of face matching. Experiments are performed using the FRGC v2. 0 dataset to demonstrate the effectiveness of our algorithm. Results confirm the expression-robustness of the proposed method.