A Study On Gender Classification Based On Facial Image Features

The main target of this thesis is to study multi-view gender classification based on facialimage. Gender classification based on facial image which is a very popular research topicin the fields of pattern classification and computer vision is a large-scale, complicated two-class classification problem by nature. Gender classification consists of three main parts: 1)face detection in the original image; 2) feature extraction from facial image; 3) pattern clas-sification. In this paper, we mainly concentrate on the face detection and feature extractionparts and especially analyze the performances of applying our proposed feature extractionmethods into the gender classification.The highlights and main contributions of the dissertation include:1. A dynamic cascade algorithm for face detectionA novel method, called"Dynamic Cascade", for training an efficient face detector isproposed in this paper. This method can train cascade classifiers on massive data sets andonly requires a small number of training parameters. Moreover, it enables efficient paralleldistributed learning. So it takes no more than 8 hours to train a detector on a training set with10 billion samples by using 30 desktop computers.2. A new weak classifier called"Bayesian Stump"for training boost classifierWe propose a new kind of weak classifier, called"Bayesian Stump", for training boostclassifiers to produce more stable boost classifiers with fewer number of features. In order tominimize the expected Bayesian error,"Bayesian Stump"extends the naive decision stumpto a single-node multi-way split decision tree method.3. Training a face detector by using multiple sets of featuresTwo challenging problems for face detection are to reduce the computational cost andimprove the detection accuracy. To address these two problems, we use multiple levelsof features sets based on the dynamic cascade structure, applying the less computationallyintensive features first. In this paper, three complementary feature sets are used: Haar-likefeatures, Gabor wavelet features, and local edge orientation histogram features. 4. A feature extraction method based on local Gabor binary mapping patternsA novel face representation approach, local Gabor binary mapping pattern, is proposedfor multi-view gender classification. In this approach, a face image is first represented asa series of Gabor magnitude pictures (GMPs) by applying multi-scale and multi-orientationGabor filters. Each GMP is then encoded as a local Gabor binary pattern (LGBP) image,where a uniform local binary pattern (LBP) operator is used. After that, each LGBP imageis divided into non-overlapping rectangular regions, from which spatial histograms are ex-tracted. To reduce the dimension of the regional LGBP features, we propose two strategiesfor mapping each of regional LGBP features onto a one-dimensional subspace independentlybefore they are concatenated as a feature vector. By using the proposed two strategies, thefeature dimension is even less than that of the features extracted by using LBP directly ongray-scale images.