不同规模数据集下的人脸识别方法(英文)

Face Recognition on Datasets of Various Scales

系统研究了类内变化和类数目增加所引起的人脸识别中的非线性识别问题,并比较了线性识别方法和非线性识别方法在不同用户集规模下的适用性.采用CAS-PEAL大型人脸数据库中的表情集(330人)和姿势集(1 000人)进行了3组实验.实验结果表明:当训练集的人数在300人(表情集)以内时,增加类内的变化不会对线性识别方法造成影响,并可以提高识别的准确率;但是,当保持类内图片数不变而增加类的数目时,类数(人数)增加对线性方法和非线性方法产生了不同的影响.随着人数增加,线性识别方法的识别准确率逐渐降低,而基于核方法的非线性方法却能够一直保持识别准确率的稳定.因此,应该根据类的总数合理地选择识别方法,并合理地设计类内的图片数,这样有助于提高人脸识别系统的识别率.同时,实验也验证了基于核方法(kernel)的非线性人脸识别方法更适宜于人数规模较大的情况.

The nonlinear recognition problem was discussed while increasing varieties within classes or numbers of classes, and the recognition results was compared by using linear and nonliner methods on the face datasets of various scales. Three experiments were carried out using the expression set（330 persons）and the pose set （1 000 persons）from CAS-PEAL face database. The results show that only increasing varieties within classes can improve correct recognition rates using either linear methods or nonlinear methods while the number of classes is invariable and smaller than 300. However, while the number of the classes increases and the varieties within classes keeps invariable, the results by using the linear recognition methods and the nonlinear methods are different. With the number of the classes increasing, the correct recognition rates decrease by using the linear methods, but for the nonlinear method based on kemel they are relatively invariable. Therefore, it is concluded that reasonably selecting methods and designing varieties within classes according to the number of classes are necessary to improve correct recognition rate. Furthermore, the nonlinear method based on kernel is fit for datasets in large scale.