基于单视图多姿态的人脸识别方法研究

发布时间：2019-06-15 14:06

【摘要】：人脸识别作为少数几个同时具有高精度和低干涉的生理特征识别方法,在数字身份认证、公共安全、多媒体等领域具有重要的应用价值。目前,在控制配合条件下的人脸识别系统能够取得较高的识别率,但当人脸存在姿态变化时,同时是单视图时,人脸识别系统面临巨大挑战。本文围绕单视图多姿态的人脸识别方法进行了系统研究,具体工作和主要成果包括:1、提出—种基于线性回归算法与支持向量机相结合的方法。针对待识别人脸存在姿态变化,基于线性回归算法寻求正、侧人脸之间的关系,然后利用此关系进行姿态校正。最后,利用支持向量机在小样本分类上的优势,采用遗传算法筛选其参数,对校正后的待识别人脸进行分类识别。在CAS-PEAL-R1人脸库上,识别率达86%。实验结果表明,该方法在处理基于单视图多姿态的人脸识别问题时,识别率高于同类其它方法。2、提出一种基于单张三维人脸重建生成虚拟多视图的方法。针对训练样本不足的问题,借助基于稀疏形变模型的三维人脸重建方法,重建输入人脸的三维人脸模型,然后通过纹理映射、旋转、投影等方法获取输入人脸的多姿态人脸图像,丰富训练样本库。在此基础上,利用BP神经网络进行人脸识别。在CAS-PEAL-R1人脸库上,识别率达91%。实验结果表明,该方法生成的虚拟多视图提高了识别效果,识别率高于同类其它方法。3、在上述工作的基础上,设计并实现了基于Matlab的人脸识别系统。该系统集成了本文所提的两种方法。
[Abstract]:Face recognition, as a few physiological feature recognition methods with high precision and low interference at the same time, has important application value in digital identity authentication, public security, multimedia and other fields. At present, the face recognition system can achieve a high recognition rate under the condition of control and coordination, but when the face posture changes and it is a single view, the face recognition system faces great challenges. In this paper, the face recognition method based on single view and multi-pose is studied systematically. the concrete work and main results are as follows: 1. A method based on linear regression algorithm and support vector machine (SVM) is proposed. Aiming at the attitude change of other people's faces, the relationship between positive and lateral faces is found based on linear regression algorithm, and then the attitude correction is carried out by using this relationship. Finally, using the advantages of support vector machine in small sample classification, genetic algorithm is used to screen its parameters, and the corrected human face is classified and recognized. On CAS-PEAL-R1 face database, the recognition rate is 86%. The experimental results show that the recognition rate of this method is higher than that of other similar methods when dealing with the problem of face recognition based on single view and multi-pose. 2, a virtual multi-view method based on single 3D face reconstruction is proposed. In order to solve the problem of insufficient training samples, the 3D face model of input face is reconstructed with the help of 3D face reconstruction method based on sparse deformation model, and then the multi-pose face image of input face is obtained by texture mapping, rotation, projection and so on, which enriches the training sample database. On this basis, BP neural network is used for face recognition. On CAS-PEAL-R1 face database, the recognition rate is 91%. The experimental results show that the virtual multi-view generated by this method improves the recognition effect, and the recognition rate is higher than that of other similar methods. 3. On the basis of the above work, a face recognition system based on Matlab is designed and implemented. The system integrates the two methods proposed in this paper.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】