【作者】 王鹏；

【导师】 穆国旺；

【作者基本信息】 河北工业大学 ， 计算数学， 2015， 硕士

【摘要】 性别识别技术涉及模式识别、心理学、人工智能、计算机视觉等不同领域,在人机对话、图像与视频检索、人口信息采集、安全门禁、身份认证等方面都有着一定程度的应用,因此被各个领域的研究人员所关注。然而,性别识别技术仍然处于探索阶段,不同的算法都存在一定的局限性,如何提高性别识别的准确性及其自动化程度依旧是一个值得研究的问题。由于利用典型相关分析(CCA)将同一模式下的两组特征矢量进行融合的方法在人脸识别、手写体识别、表情识别等问题中都体现出其优越性,因此,本文重点研究了CCA在性别识别中的应用。本文的主要工作如下:1.对性别识别中常用的一些特征提取和分类方法进行了介绍,具体包括局部二值模式(LBP)、梯度方向直方图(HOG)和C1等特征提取方法,以及支持向量机(SVM)、AdaBoost和随机森林(RF)等分类决策方法。使用这些方法在MIT和VIPeR行人图像数据库上进行了性别分类实验,对不同的特征提取和分类方法进行了比较研究。实验结果表明,在性别识别中,C1特征优于LBP特征和HOG特征,利用SVM进行性别识别具有较好的性能。2.提出了基于CCA特征融合和SVM的全身图像性别识别方法。首先,利用CCA方法对LBP特征、HOG特征以及C1特征进行两两融合,得到LBP-HOG、LBP-C1和HOG-C1三种组合特征,然后,使用SVM方法进行性别识别。在MIT和VIPeR图像数据库上进行了实验,对三种组合特征以及单一特征在性别识别中的优劣进行了比较。实验结果表明,在性别识别中,利用CCA融合得到的组合特征优于单特征。在三种组合特征中,HOG-C1特征识别效果最好。3.研究了朝向对全身图像性别识别的影响,提出了一种基于CCA特征融合、并考虑朝向的全身图像性别识别算法。首先,利用CCA方法对HOG和C1特征进行融合,然后,利用SVM对朝向进行分类。在预测朝向的过程中,根据测试样本到SVM超平面距离将朝向归为三种情况,即正面朝向、背面朝向以及正面与背面混合朝向。然后根据每一种情况进一步进行性别识别。实验结果表明,本文提出的方法使得性别识别率进一步提高。更多还原

【Abstract】 Gender recognition is involved in pattern recognition, psychology, artificial intelligence,computer vision, it has a certain degree of applications in human-machine dialogue, image and video retrieval, demographic information collection, security access system, identity authentication,etc,therefore it has received widely attentions. However, gender recognition is still in an exploratory stage,and each method has its limitations. How to improve automaticity and accuracy of gender identification is still a problem worthy of studying. Because the method that use canonical correlation analysis(CCA) to fuse two groups of features of one pattern has shows great superiority in face recognition, handwriting recognition and expression recognition,the application of CCA in gender recognition is focused on in this paper. The main contribution is as follows:1. Some widely used methods for feature extraction such as local binary pattern(LBP),histogram of oriented gradient(HOG), C1 features and the methods for classification including support vector machine(SVM), AdaBoost and random forest(RF) are described. Experiments on gender recognition using these methods are conducted on MIT and VIPeR databases of pedestrian images, different features and classification methods are compared. The experimental results show that C1 features is better than LBP and HOG features, and using SVM can obtain better performance in gender recognition.2. An algorithm for gender recognition from body images based on canonical correlation analysis(CCA) and SVM is presented. First, LBP, HOG and C1 features are pairwise fused using CCA method to obtain three kinds for fused features including LBP-HOG,LBP-C1 and HOG-C1 feature, then the fusion features are inputted into SVM for gender recognition. The three kinds of original features and three kinds of fused features are compared by experiments on MIT and VIPeR database. The experimental results show that fusion feature by CCA is better than single feature,moreover,HOG-C1 feature works best.3. The influence of orientation on gender recognition is considered,and a new algorithm for gender recognition from body image based on CCA and the orientation classification is presented.First,HOG and C1 features are fused using CCA method,then SVM is used for orientation classification. According to the distance from test sample to the SVM hyperplane, the orientation of images are grouped into three classes, i.e., front view,mixed view and back view. Finally, one classifier trained for the view of the test image is used for gender recognition. The experimental results show that the proposed method makes further improvement in recognition rate of gender recognition.更多还原