Pseudo-Example Based Iterative SVM Learning Approach for Gender Classification

Abstract

In order to increase the detection accuracy in gender classification, a pseudo-example based iterative learning approach combining support vector machine (SVM) and active appearance model (AAM) was proposed. AAM was applied to model the original training examples before constructing the SVM classifier. During the current iteration, some pairs of support vectors with different gender were selected randomly and then their AAM parameters were interpolated properly to generate new pseudo face images as candidate examples with new gender feature pattern. Only the candidates that would be classified by the current classifier incorrectly or correctly but with low confidence were selected for the following iterations. The pseudo-examples created in this way complemented the original training examples effectively, and the proposed pseudo-example selecting scheme outperformed the conventional Bootstrap method. Experimental results show that, this iterative learning approach can upgrade the gender detection accuracy stepwise