电子学报 ›› 2017, Vol. 45 ›› Issue (6): 1462-1468.DOI: 10.3969/j.issn.0372-2112.2017.06.026

• 学术论文 • 上一篇    下一篇

基于混合属性的零样本图像分类

程玉虎, 乔雪, 王雪松   

  1. 中国矿业大学信息与控制工程学院, 江苏徐州, 221116
  • 收稿日期:2015-05-13 修回日期:2015-12-12 出版日期:2017-06-25
    • 通讯作者:
    • 王雪松
    • 作者简介:
    • 程玉虎 男,1973年出生于安徽淮南,中国矿业大学信息与控制工程学院教授,博士生导师.主要研究方向为机器学习与模式识别、智能系统等.E-mail:chengyuhu@163.com;乔雪 女,1990年出生于四川成都,中国矿业大学信息与控制工程学院硕士研究生.主要研究方向为迁移学习、属性学习等.E-mail:qiaoxuecumt@163.com
    • 基金资助:
    • 国家自然科学基金 (No.61472424,No.61273143); 中央高校基本科研业务费 (No.2013RC10,No.2013RC12,No.2014YC07)

Hybrid Attribute-Based Zero-Shot Image Classification

CHENG Yu-hu, QIAO Xue, WANG Xue-song   

  1. School of Information and Control Engineering, China University of Mining & Technology, Xuzhou, Jiangsu 221116, China
  • Received:2015-05-13 Revised:2015-12-12 Online:2017-06-25 Published:2017-06-25
    • Supported by:
    • National Natural Science Foundation of China (No.61472424, No.61273143); Fundamental Research Funds for the Central Universities (No.2013RC10, No.2013RC12, No.2014YC07)

摘要:

对于具有相似属性的类别而言,在有限维度的语义属性下,基于属性的零样本图像分类器难以对它们进行正确区分.考虑到语义属性描述类别的有限性,在直接属性预测(Direct Attribute Prediction,DAP)模型的基础上,提出一种基于混合属性的零样本图像分类模型(Hybrid Attribute-Based DAP,HA-DAP).首先,对样本的底层特征进行稀疏编码并利用编码后的非语义属性来辅助现有的语义属性;将非语义属性与语义属性构成混合属性并将其作为DAP模型的属性中间层,利用属性预测模型的思想进行混合属性分类器的训练;最后,根据预测的混合属性以及属性与类别之间的关系进行测试样本类别标签的预测.在OSR、Pub Fig以及Shoes数据集上的实验结果表明,HA-DAP的分类性能优于DAP,不仅能够取得较高的零样本图像分类精度,而且还获得了较高的AUC值.

关键词: 零样本图像分类, 混合属性, 语义属性, 非语义属性, 稀疏编码

Abstract:

When the dimensionality of the semantic attributes is limited,it is difficult for attribute-based zero-shot image classifiers to distinguish the objects with similar attributes.Aiming at the limitation of describing objects with semantic attributes,an improved direct attribute prediction(DAP) model for zero-shot image classifying based on hybrid attribute(HA) is proposed,which is called HA-DAP.At first,we carry out the sparse coding on the low-level features to obtain the non-semantic attributes that are used to assist the existing semantic attributes.Then,we take the hybrid attributes including the learned non-semantic attributes and the manually specified semantic attributes as the mid-layer of DAP model and use the idea of attribute prediction to train the hybrid attribute-based classifier.At last,according to the predicted hybrid attributes and the relationship between the attributes and classes,we can recognize the class label for the testing sample.Experimental results on the OSR,Pub Fig and Shoes datasets show that,the HA-DAP outperforms the DAP in the classification performance,i.e.,when compared with the DAP,the proposed HA-DAP yields much higher zero-shot image classification accuracy and AUC value.

Key words: zero-shot image classification, hybrid attribute, semantic attribute, non-semantic attribute, sparse auto-encoding

中图分类号: