Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (3): 524-532.DOI: 10.3778/j.issn.1673-9418.2006071

• Graphics and Image • Previous Articles     Next Articles

Facial Image Based Two-Level Model for Gender Classification

YANG Chenxu, CAI Kecan, ZHANG Hongyun, MIAO Duoqian   

  1. Department of Computer Science and Technology, Tongji University, Shanghai 201804, China
  • Online:2021-03-01 Published:2021-03-05

基于人脸图像的二阶段性别分类算法

杨晨旭蔡克参张红云苗夺谦   

  1. 同济大学 计算机科学与技术系,上海 201804

Abstract:

Many scenes need facial gender identification with good accuracy. Deep convolutional neural networks (CNN) with large set of training data normally give good accuracy, however, to achieve good accuracy with uncertain training data is a difficult task due to their lower explanation and potential information loss. Moreover, the uncertain-ties resulted from illumination, postures and facial expressions can lead to low accuracy of the classification. In this paper, a shadowed sets based two-level model for gender classification is proposed to address the problem. Deep convolutional neural networks are used as one-level classifier. Combining the concept of shadowed sets, one-level classification results are divided into three parts: accept domain, reject domain and uncertain domain. Samples in the uncertain domain are selected as uncertain facial images for two-level reclassification. Results show that the proposed method can further improve the accuracy compared with several existing state-of-the-art methods on the LFW dataset and Adience dataset.

Key words: gender identification, convolutional neural networks (CNN), shadowed sets, uncertain domain

摘要:

许多现实场景要求准确的脸部性别识别。深度卷积神经网络在正常状况下取得好的准确率,适用于大规模分类任务,但存在模型可解释性差、易丢失细节信息等问题,并且光照、姿势、表情等因素带来的不确定性会导致分类准确率较低。提出一种基于阴影集的二级分类模型。采用深度卷积神经网络对大规模图像集进行一阶段分类;结合阴影集理论,将图像分类结果划分为接收域、拒绝域和不确定域,得到不确定的脸部图像集,用传统方法进行二阶段分类。在LFW数据集和Adience数据集下,与现有先进算法相比,所提方法能有效地提高总体分类的准确率。

关键词: 性别识别, 卷积神经网络(CNN), 阴影集, 不确定域