Face detection using kernel PCA and imbalanced SVM

The task of face detection can be accomplished by performing a sequence of binary classification: face/nonface classification, in an image. Support vector machine (SVM) has shown to be successful in this task due to its excellent generalization ability. However, we find that the performance of SVM is actually limited in such a task due to the imbalanced face/nonface data structure: the face training images outnumbered by the nonface images in general, which causes the class-boundary-skew (CBS) problem. The CBS problem would greatly increase the false negatives, and result in an unsatisfactory face detection rate. This paper proposes the imbalanced SVM (ISVM), a variant of SVM, to deal with this problem. To enhance the detection rate and speed, the kernel principal component analysis (KPCA) is used for the representation and reduction of input dimensionality. Experimental results carried out on CYCU multiview face database show that the proposed system (KPCA+ISVM) outperforms SVM. Also, results indicate that without using KPCA as the feature extractor, ISVM is also superior to SVM in terms of multiview face detection rate.