Improving generalization for gender classification
Xue-Ming Leng, Yiding Wang · 2008
This paper addresses the problem of improving the generalization ability for gender classification. An approach based on Fuzzy SVM (FSVM) is developed to improve it. The fuzzy membership used in FSVM indicates the degree of one person’s face belonging to female/male faces. Based on Learning Vector Quantization (LVQ) learning process, a novel method of generating fuzzy membership function automatically is proposed in this paper. The method doesn’t rely on the apriori information of data and generates the membership function as objective as may be. The gender classifier based on FSVM is evaluated on the FERET, CAS-PEAL, BUAA-IRIP face databases. The results show that the gender classifier presented in this paper can tolerate more variations such as illumination, expression and pose and show good performance in generalization ability.