A Study Of Gender Classification Based On Images

The biological characteristics are the inherent attributes of human beings, which havestrong self-stability and individual independency. Therefore, it has been a common researchproblem to apply the individual biological characteristics into machine learning intelligenceand make the computer more intelligent as human beings. Recently, several different bio-logical characteristics have been applied into pattern recognition including face recognition,fingerprint recognition, iris recognition, palm print recognition and etc. As a representative,face recognition has become popular and been improved a lot. Gender classification is a sub-problem of face recognition which can be used for not only business statistical applicationsand robotics but also for evaluation about performance of large-scale classifiers. Similar asface recognition, gender classification consists of three main parts: image preprocessing,feature extraction and pattern classification. In this paper, we mainly focus on the featureextraction and pattern classification parts and especially analyzed how to use hair feature toclassify the gender.Image preprocessing mainly concentrate on image grayed, image geometry normaliza-tion and intensity histogram equalization. These will complete the image normalization andhelp improve the image quality, decrease the computation complexity and therefore enhancethe recognition accuracy and accelerate the convergence speed.In the image feature extraction step, we brie?y review the principle and usage of theclassical methods: Eigen-face, Fisher-face and Gabor wavelet feature. Then we propose ahair detection method on the front view face image and define 6 different hair features aswell as the corresponding methods to calculate them.In the gender classification step, our main effort are about using Min-Max ModularSVM(M3-SVM) to discriminate different gender. M3-SVM is composed of two parts: taskdecomposition and sub-problem combination. This Min-Max Modular scheme is based on a'divide and conquer'strategy and it can divide a large-scale problem into several independentsmaller sub-problems and then combine the results of each sub-problem into a solution of theoriginal problem. Firstly, we analyzed the theory of Min-Max Modular network and its com-bination with SVM. Then we introduce 3 task decomposition strategies including randomlydecomposition, decomposition with hyper-plane and decomposition with prior knowledge. At last, we present the experiment result on 3 different face databases.Firstly, we com-pare the representation abilities of different feature extraction methods and the result showsthat the hair feature outperformed other related methods. Secondly, we compare the per-formance of M~3-SVM using different task decomposition strategies with SVM. The resultshows that the M~3-SVM can improve both the accuracy and computation speed for genderclassification.