Research On Expression-Iinsensitive Three Dimensional Face Recognition

Face recognition is one of the most popular and challenging subjects of biological feature identification technology. As the most unconstraint and friendly authentication method, face recognition becomes one of the most promising technologies in this century. With the rapid development of numerous related fields, such as image processing, pattern recognition, machine learning and computer vision, face recognition has been in great need of public safety and national security, and becomes one of the most important technique methods in that field. Although 2D face recognition technology has a history of more than ten years, it still bears limitations which restrict the application of face recognition technology in public safety. Especially in the cases of make-up, pose and expression changes, as well as the change of illumination, the accuracy of the 2D face recognition algorithm is significantly decreased.With the rapid development of 3D data acquisition equipment, more and more researchers have committed to use three dimensional face features for recognition and have gain good results. At the same time, three dimensional face recognition methods promote the development of face recognition. However, facial expression changes are still one of the biggest challenges of 3D face recognition. This paper focuses on 3D facial expression changes, and presents expression-insensitive features for recognition. The main propose is extracting similar feature from different expressional faces of same person and discriminant feature from different person, and laying a solid foundation for accurate and efficient face recognition. Based on the detailed analysis of 3D face data, deeply study and discussion on the key problems of the restriction for 3D face recognition, this paper proposes a 3D face recognition system containing 3D data prepossessing, expression-insensitive feature extracting and 3D face recognition method, and achieves good results. The main researches and innovative works are as follows:1. Propose an iterative nose tip detection method based on self-symmetryThis paper proposes an iterative nose tip detection method based on self-symmetry. The strategy of the nose detection method is gradually narrowing the scope of the nose region. First, the facial symmetrical plane is obtained based on the symmetry of 3D face, then a standard central stripe is aligned to the intersection of symmetrical plane and 3D face. The nose tip is obtained as the highest point in Z direction in the region which centered at the nose tip of standard 3D face. Finally, for adjusting the nose tip, the spherical region which centered at the nose tip is alighed to the standard face. The new nose tip is obtained using the highest point in the nose region of the last step. The nose tip is adjusted iteratively until the Euclidean distance between the current nose tip and the previous one is less than a threshold. This method could deal with expressional faces, faces which contain hair, and faces with small angle. And it is helpful for feature extraction and classification algorithm subsequently. Finally, the face region which is extracted in the last step is used for recognition.2. Propose a neutral face estimation map based on deformation modelIn order to reduce the effect of expression of 3D face recognition result, this paper proposes a L1 norm constraining expressional space coefficients least squares regression algorithm. For a testing face, the method first establishes the neutral space and expressional space, and then uses a deformation model to approximate the 3-D surface shape. Because people makes one expression in a specific period of time, so the coefficients of the expressional space of the deformation model are constrained by L1 norm, then the parameters of model are estimated by L1 norm constraining expressional space coefficients least squares regression algorithm. The coefficients of neutral space and expressional space are used to estimate the testing face subsequently. Finally, in order to maintain the detailed information of the original face, this method uses sum of the neutral component and the standard face to establish the corresponding relationship with the original face, generating a neutral face estimation map. Experiments show that using neutral face estimation map for feature extraction could increase the similarity between expressional 3D face with its corresponding neutral face, thereby reducing the within class scatter and the effect of expression on the face recognition algorithm, and finally improving the recognition rate.3. Propose spherical vector norms map based on norm vectorIn order to effectively reflect the shape of the face surface, this paper presents a spherical vector norms map (SVNs map) using norm vector of cloud data. The SVNs map reflects the characteristics of 3D face on a sphere. The spherical vector is defined by a vector that originates from a point on the 3D face and ends on a spherical surface, and then each point’s norm of spherical vector (SVN) is obtained. Subsequently, SVN substitutes each point’s z value for increasing the distinguishability of different points. In addition, for the mouth region, the algorithm adjusted the spherical vector norms of the mouth using a standard face. This makes spherical vector norms map insensitive to facial expression. Therefore using spherical vector norms maps as feature can not only improve the similarity between expressional 3D face and neutral face of same person, but also decrease the similarity between 3D faces which have the same expression from different classes. Thereby the map could reduce the within-class scatter, and enhance the between-class scatter, and finally improve the recognition rate of 3D face recognition.4. Propose a 3D face feature based a new partitioning strategy of Histograms of Oriented GradientsIn order to extract local feature of 3D face and maintain the integrity of the feature region of the face, a new partitioning method is adopted to extract the histograms of oriented gradients (New Partitioning Histograms of Oriented Gradients, NP-HOG) of the neutral face estimation map and the spherical vector norms map. The projection map is divided into eight blocks of different sizes. Each block’s histograms of oriented gradients feature is connected to a vector as a NP-HOG feature. This paper also proposes a fusion method for 3D face recognition algorithm based on NP-HOG feature. Firstly, the neutral face estimation map and the spherical vector norms maps’NP-HOG features are extracted. Then the two NP-HOG features are projected into the LDA subspace respectively. Cosine distance is applied to establish two similarity matrices. Sum rule is used to fuse the two similarity matrices and the Nearest Neighbor classifier is applied to finish the recognition process. The NP-HOG features are extracted from the expression-insensitive projection maps, which can reduce the impact of the facial expression on the recognition stage, and increase the between-class scatter. The experimental results also show that the NP-HOG feature is more appropriate for 3D face recognition than some commonly used facial features, and the fusion algorithm can further improve the recognition rate.