Computer Engineering and Applications ›› 2016, Vol. 52 ›› Issue (14): 136-141.

Previous Articles     Next Articles

Multi-scale unsupervised feature learning for face recognition

YIN Xiaoyan1, FENG Zhiyong1, XU Chao2   

  1. 1.School of Computer Science and Technology, Tianjin University, Institute of Knowledge Science and Engineering, Tianjin 300072, China
    2.School of Computer Software, Tianjin University, Tianjin 300072, China
  • Online:2016-07-15 Published:2016-07-18

多尺度非监督特征学习的人脸识别

尹晓燕1,冯志勇1,徐  超2   

  1. 1.天津大学 计算机科学与技术学院,知识科学与工程研究所,天津 300072
    2.天津大学 软件学院,天津 300072

Abstract: In order to fully utilize latent information of human face, a method called Multi-Scale Convolutional Auto-
Encoder(MSCAE) is proposed. MSCAE extracts image’s multi-scale features using different sizes of convolution kernels. Since the new features reflect natural facial contents, human face can be restored better. The MSCAE applies a hierarchy of alternating filtering and sub sampling, and it makes features invariant to deformations including rotation, translation, and scale. The form of encoder-decoder is introduced to train the MSCAE so as to obtain the feature extractor and vectors combining multi-scale features for further classification. Experiments are conducted with Neural Network(NN) on ORL and Yale face datasets, and the experimental results suggest that multi-scale features are superior to single-scale ones on recognition rate and efficiency. Furthermore, fusion features of MSCAE and Histograms of Oriented Gradients(HOG) can get higher recognition rate than either of them.

Key words: unsupervised feature learning, multi-scale, convolutional auto-encoder, deep learning

摘要: 为了充分利用人脸图像的潜在信息,提出一种通过设置不同尺寸的卷积核来得到图像多尺度特征的方法,多尺度卷积自动编码器(Multi-Scale Convolutional Auto-Encoder,MSCAE)。该结构所提取的不同尺度特征反映人脸的本质信息,可以更好地还原人脸图像。这种特征提取框架是一个卷积和采样交替的层级结构,使得特征对旋转、平移、比例缩放等具有高度不变性。MSCAE以encoder-decoder模式训练得到特征提取器,用它提取特征,并融合形成用于分类的特征向量。BP神经网络在ORL和Yale人脸库上的分类结果表明,多尺度特征在识别率和性能上均优于单尺度特征。此外,MSCAE特征与HOG(Histograms of Oriented Gradients)的融合特征取得了比单一特征更高的识别率。

关键词: 非监督特征学习, 多尺度, 卷积自动编码器, 深度学习