The influence of ethnicity in facial gender estimation
The influence of ethnicity in facial gender estimation
复制标题
种族对面部性别估计的影响
DOI：
10.1109/cspa.2018.8368710
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
C. H. Lim
中科院分区：
文献类型：
--
作者：
E. K. Loo;T. S. Lim;L. Ong;C. H. Lim
面部性别估计现在非常有名。它已被广泛应用于生物识别、定向广告和人机交互等领域。马来西亚是一个开始利用上述技术的发达国家。直到今天，已经开发了大量的面部性别估计算法。据我们所知，还没有研究人员用马来西亚民族数据库发布过现有算法的基准。此外，没有公开的马来西亚族裔数据库。因此，现有算法对马来西亚的性能是未知的。除此之外，不同的种族在外表上也有差异。马来西亚是一个多种族的国家，由多个种族组成。因此，马来西亚民族可能看起来与其他民族不同，如西方人的计算机观点。本文提供了一个实验研究，以显示现有的算法对马来西亚民族的性别估计性能。最重要的是找出种族在性别估计中的影响。为了了解民族的影响，面部识别技术（FERET）数据库和马来西亚民族面部数据库（MEFD）被用来代表非马来西亚民族和马来西亚民族。选择主成分分析（PCA）和多层次局部二值模式（ML-LBP）两种特征提取方法，分别以支持向量机（SVM）作为分类器进行实验。所有的实验都是在没有人脸对齐的情况下进行的。结果表明，ML-LBP取得了更好的准确性比主成分分析在马来西亚种族和非马来西亚种族的性别估计。最后但并非最不重要的是，种族确实影响性别估计。训练集必须涉及目标的种族，以便具有良好的准确性。
Facial gender estimation is very famous nowadays. It has been widely applied in many applications such as biometrics, targeted advertising and human-computer interaction. Malaysia is a developed country that begins to utilize the mentioned technology. Until today, there are plenty of algorithms for facial gender estimation have been developed. For our best knowledge, no researcher ever publish a benchmark of existing algorithms with Malaysian ethnics' database. Moreover, there is no publicly available Malaysia ethnics' database. Hence, the performance of the existing algorithms towards Malaysian is unknown. Other than that, different ethnics have differences in physical look. Malaysia is a multiracial country that consists of multiple races. Thus, Malaysian ethnics may look different with other ethnics such as westerners for computer point of view. The paper provides an experimental study to shows the gender estimation performance of the existing algorithms towards Malaysian ethnics. Most importantly is to find out the influence of ethnicity in gender estimation. To find out the influences of ethnics, Facial Recognition Technology (FERET) Database and Malaysian Ethnics Facial Database (MEFD) are used to represent non-Malaysian ethnics and Malaysian ethnics. Two feature extraction methods namely Principle Component Analysis (PCA) and Multi-Level Local Binary Patterns (ML-LBP) are chosen to conduct the experiments independently with Support Vector Machine (SVM) as a classifier. All the experiments are conducted without face alignment. The result shows that ML-LBP achieved better accuracy than PCA in gender estimation for both Malaysian ethnics and non-Malaysian ethnics. Last but not least, the ethnicity does affect the gender estimation. The training set has to involve the target's ethnicity in order to have a good accuracy.