局部特征脸型分类方法

PDF下载HTML

Title:

Local feature facial classification method

作者:

孙劲光, 邓智硕

辽宁工程技术大学 电子与信息工程学院, 辽宁 葫芦岛 125105

Author(s):

SUN Jinguang, DENG Zhishuo

School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China

关键词:

脸型分类; 圆形邻域; 特征编码; 局部特征表达; 多分类支持向量机

Keywords:

facial classification; round-neighborhood; feature coding; local feature representation; multi classification support vector machineface

分类号:

TP391.41

DOI:

10.11992/tis.201605021

摘要:

本文针对传统脸型分类算法特征点定位不准和过度依赖轮廓曲线的问题，提出了一种人脸轮廓圆形邻域局部特征表达方式和脸型分类模型。首先，初步定位脸型轮廓特征点；然后，在特征点周围选取三重八连通圆形邻域，通过计算一级邻域、拓展邻域与中心区域间的纹理变化，生成二进制编码序列，构造脸型局部特征向量；最后，设计OVO-RBF-SVM多分类模型，实现脸型分类。本文方法在CAS-PEAL人脸库上进行脸型类型判别，获得了94.28%的准确率；在相同情况下，分别与基于主动形状模型和基于下颌曲线模型的脸型类型判别方法进行对比，准确率分别提高了6.64%和6.58%。本文所研究的方法在一定程度上解决了特征点定位相对不准确导致误差增加的问题，同时尽可能多利用图片原始信息，保证轮廓特征提取的准确率，具有较强的鲁棒性。通过实验证明本文方法适用于脸型分类。

Abstract:

Considering the problems where the feature points of traditional facial classification algorithms are not located in the position of the actual feature points and are heavily dependent upon the contour curve, a facial contour circular neighborhood local feature expression and a facial classification model were proposed.First, the preliminary facial contour feature points were located and then around the feature points, the triple eight connected round-neighborhood was selected.By calculating a neighborhood level and expanding the neighborhood with the central area between the texture changes, the binary code sequence was generated and the tectonic facial local feature vectors can be created. Then, the faces were classified by designing the OVO-RBF-SVM classification model. The experiment was conducted on the CAS-PEAL face library for facial contour feature discrimination, achieving 94.28% accuracy rate; under the same circumstances, the face-type discrimination methods which are based on the active shape model and jaw curve model were compared, and the accuracy rate raised 6.64% and 6.58%, respectively. To a certain extent, the method proposed in this paper solves the problem where the error increases when the location of the feature points are relatively inaccurate, and at the same time, the original picture information is utilized as much as possible, to ensure the accuracy of the contour feature extraction, which has strong robustness. The experimental results show that this method is suitable for facial classification.