Font Size: a A A

The Research Of Gender Classification And Classifier Confidence

Posted on:2012-04-18Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z JiFull Text:PDF
GTID:1118330362958331Subject:Computer software and theory
Abstract/Summary:PDF Full Text Request
Years of pattern recognition studies have shown that automatic gender classification, as an important part of computer vision, is attractive due to its wide range of application such as video surveillance, robot vision, intelligent human-machine communication as well as kee-jerk demographic data collection. This thesis focuses on the research of gender classification based on image feature and classifier confidence. The main contributions of this thesis exist in the following aspects:1) we propose a face feature extraction method, multi-resolution local Gabor binary pat-tern (MLGBP), for gender classification. The idea is to combine the characteristics of multi-resolution analysis with Gaussian pyramid, Gabor filters and local binary pattern. The main contributions of MLGBP are twofold as follow:(a) Both fine and coarse local micro-patterns and spatial information of a series of facial image obtained by multi-resolution analysis with Gaussian pyramid are maintained; (b) We propose two approaches to segment those Gabor images. One approach is that each Gabor image is equably divided into non-overlapping rectangular regions, and the other is that each Gabor image is proportionally partitioned according to its size. To reduce the number of feature dimensions, two ways have been pro-posed. One is linear discriminant analysis (LDA) making MLGBP-LDA and the other is to project the regional features onto the class center connecting line (CCL) making MLGBP-CCL. The experimental results show that in most cases MLGBP-CCL is superior to all of the six existing methods.2) A method of gender classification fusing hair and face features are presented. Firstly, we obtain a geometric hair model (GHM) by means of binary filter, edge recognition, contour extraction, and wavelet transformation and use them to extract some feature attributes such as length, area, color, texture, and split location on hair region. Secondly, local binary pattern (LBP) algorithm is utilized to extract face feature after normalizing and preprocessing facial images. In order to integrate information from a variety of classifiers, we propose a fuzzy integral fusion model based on support vector machine (FIF-SVM). The experiments are done on AR, CAS-PEAL and FERET databases and experimental results show that FIF-SVM can improve classification rate by making use of info fusion than only single feature.3) We explore how to incorporate human prior knowledge into machine learning and propose a support vector machine with confidence (SVMC), which is introduced in the prim-itive training samples. The main contributions of the SVMC method are twofold. One is that we derive the quadratic programming problem for SVMC and analyze whether SVMC can improve classification performance in theory in some kind of situations, and the other is that we compare SVMC with traditional support vector machines on a gender classification prob-lem. Experimental results demonstrate that SVMC can improve the classification accuracy dramatically when the confidence values of all the training samples are labeled appropriately.4) A support vector machine with automatic confidence (SVMAC) for pattern classi-fication is proposed. The main contribution of SVMAC to learning machines is that we develop an algorithm for calculating the label confidence value of each training sample. Thus, the label confidence values of all of the training samples can be considered in training support vector machines. To demonstrate the effectiveness of the proposed SVMAC, a series of experiments are performed on three benchmarking pattern classification problems and a challenging gender classification problem. Experimental results show that the generalization performance of our SVMAC is superior to that of traditional SVM.5) Embedding SVMAC classifier into min-max modular (M3) networks, we propose a min-max modular support vector machine with automatic confidence (M3-SVMAC) for gender classification based on facial images. It is known to us that on the one hand M3 networks have many advantages in the aspect of training, imbalance and parallel computing on large-scale samples. On the other hand, the generalization performance of SVMAC is superior to that of traditional SVM. Experimental results on gender classification based on facial images demonstrate that M3-SVMAC can improve classification accuracy dramatically, compared to M3-SVM.
Keywords/Search Tags:Feature extraction, Multi-resolution local Gabor binary pattern, Support vector machine, Gender classification, Fuzzy integral fusion, Classifier confidence, Min-max modular networks
PDF Full Text Request
Related items