Computer Engineering and Applications ›› 2018, Vol. 54 ›› Issue (16): 135-139.DOI: 10.3778/j.issn.1002-8331.1704-0460

Previous Articles     Next Articles

Age and gender classification using improved convolutional neural networks

CHEN Jinan1, LI Shaobo1,2, GAO Zong1, LI Zhengjie1, YANG Jing1   

  1. 1.Key Laboratory of Advanced Manufacturing Technology of Ministry of Education, Guizhou University, Guiyang 550025, China
    2.School of Mechanical Engineering, Guizhou University, Guiyang 550025, China
  • Online:2018-08-15 Published:2018-08-09


陈济楠1,李少波1,2,高  宗1,李政杰1,杨  静1   

  1. 1.贵州大学 现代制造技术教育部重点实验室,贵阳 550025
    2.贵州大学 机械工程学院,贵阳 550025

Abstract: Automatic age and gender classification of face images is an important task for face analysis. However, there is currently no model of accurate age and gender classification for face images in real world. A deep Convolutional Neural Network(CNN) is proposed to complete the recognition task. The first layer of large-scale convolution filters is replaced by cascade 3×3 convolution filters. The last hidden layer is fully connected to both the average pooling and the fifth convolutional layers such that it sees multi-scale features(features in the fifth convolutional layer are more global than those in the average pooling layer). Because of the use of Batch Normalization layer, a higher learning rate and a smaller Dropout ratio can be set. 1×1 convolution filters and global average pooling over feature maps are utilized instead of the fully convolution. The results show that the convolutional neural network proposed in this paper is superior to other state-of-the-art classification methods and achieves 89.8% accuracy on age classification, 93.3% accuracy on gender classification on Adience dataset.

Key words: deep learning, convolutional neural network, age classification, gender detection

摘要: 人脸图像的年龄和性别识别是人脸分析的重要任务,在真实多变场景下完成识别依然面临挑战。改进深度卷积神经网络(Convolutional Neural Network,CNN),将首层大尺寸卷积核替换为级联3[×]3卷积核;采用跨连卷积层融合中层和高层抽象特征;加入Batch Normalization(BN)层,设置较高的学习率和较小的Dropout比率;采用1[×]1卷积核与全局平均池化(Global Average Pooling)取代全连接层。实验表明,所提方法与主流的年龄性别识别方法比较具有较好的识别率,在Adience数据集上,年龄识别精度达到89.8%,性别识别精度达到93.3%。

关键词: 深度学习, 卷积神经网络, 年龄分类, 性别识别