The Key Techniques In 3D Face Recognition Under Non-specific Circumstances

Face recognition based on two-dimensional(2D)image has produced a large number of research results.The recognition rate achieved by computer under specific circumstances in some work even exceeds that obtained by human vision.But due to the inherent defects of 2D image,its applications are severely restricted because of the vulnerability caused by the pose rotations,illumination variations,expression changes and facial make-up,which has become a bottleneck of face recognition research.Compared to the 2D images,three-dimensional(3D)data can provide a more direct representation for a real object in 3D space.The geometrical information from 3D faces has the potential to overcome the limitations of 2D faces and achieve a greater recognition accuracy.Furthermore,the rapid development and the popularization of 3D laser scanning technology have advantageously guaranteed the acquisition of 3D data.However,because of similar structures between different subjects,variantions in 3D facial surfaces caused by changes in poses and expressions degrade recognition performance.Therefore,to be useful in real-world applications,this paper focuses on the 3D face recognition under non-specific circumstances,i.e.under the conditions of pose and expression changes,and carries out in-depth study.The achievements of the paper can be summarized as the following:(1)Based on the high-order curvature information of surface,a 3D face representation method based on ridge and ravine points is firstly applied to face recognition.As the ridge and ravine points on the 3D local surface lie on the positions that the principal curvatures change most sharply along the corresponding principal directions,the model consisted of the ridge and ravine points can represent the main features of 3D face completely and accurately.The experiments on 3D faces with poses and expressions show that the face representation based on ridge and ravine points has the advantages of high dispersion between different classes and good stability under various poses and expressions.(2)A multi-pose precise nose tip detection algorithm on 3D faces is proposed.According to the remarkable characteristics of the nose shape and the distribution characteristics of Gaussian curvature around the nose tip,the algorithm locates the nose tip by combining high order with low order curvature features on 3D faces.Moreover,it does not require training and does not depend on any particular model.The experiments on the GavabDB verify that the proposed algorithm achieve an ecouraging performance under various poses,including the poses with large rotation,such as the pure profile pose.(3)A novel 3D face pose estimation algorithm combining point feature with line feature is proposed.Based on the rotation-invariant detection results of the nose tip and nose bridge,it can get the continuous and precise pose estimation for six degrees of freedom on 3D faces under non-specific circumstances by using the accurate point feature and the stable line feature.Due to only relying on the features around the nose,which is the most prominent convex region on the whole face and is insensitive to expression variety,our algorithm is suitable for the faces with self-occlusion caused by pose change and the faces with expressions.The experimental results on the 3D synthetical faces and real faces both verify the effectiveness of our method.(4)Two methods for 3D face recognition under non-specific circumstances are proposed.In the first 3D face recognition method based on the spatial distribution of the ridge and ravine points and the Hausdorff distance,two 3D faces are firstly matched coarsely according to the histogram of spatial distribution density after creating the standard models of ridge and ravine points.Secondly,two 3D faces are matched finely according to the partial weighted LTS-Hausdorff distance.The second method for 3D face recognition is based on the fusion of multiple subject-specific curve features on the whole facial surface.These curve features are complementary to each other and can represent the important features on the whole 3D face accurately.In addition,all features are less affected by the expression.The experiments on 3D face database shows that the two 3D face recognition methods both achieve effective performance under nonspecific circumstances,especially make great improvement in the recognition rate for the 3D faces with pose and expression changes.In summary,we carry out the research on the four key technologies for 3D face recognition under non-specific circumstances in this paper,and put forward some new ideas and methods.The 3D face representation method based on ridge and ravine points provides an effective approach of feature extraction for 3D face recognition under non-specific circumstances.The multi-pose precise nose tip detection algorithm provides the important data required for key point extraction and facial model registration under non-specific circumstances.The 3D face pose estimation algorithm combining point feature with line feature provides an effective way to the pose normalization processing of 3D facial model under non-specific circumstances.Two novel 3D face recognition algorithms improve the recognition performance on the faces with poses and expressions under non-specific circumstances.