基于混合属性的零样本图像分类

发布时间：2018-05-23 13:40

本文选题：零样本图像分类 + 混合属性　；参考：《电子学报》2017年06期

【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.
[Abstract]:For the category with similar properties , it is difficult to distinguish them correctly under the semantic attributes of finite dimensions . Considering the limitation of semantic attribute description categories , a hybrid Attribute - Based DAP ( HA - DAP ) is proposed based on Direct Attribute Prediction ( DAP ) model . firstly , sparse coding is performed on the bottom layer characteristics of the sample and the encoded non - semantic attribute is utilized to assist the existing semantic attribute ; and the non - semantic attribute and the semantic attribute are used as the attribute intermediate layer of the DAP model , and the prediction of the sample category label is carried out according to the relationship between the attribute and the category ; and finally , the classification performance of the HA - DAP is better than the DAP , so that the classification accuracy of the high zero sample image can be obtained , and the higher AUC value is obtained .
【作者单位】：中国矿业大学信息与控制工程学院;
【基金】：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)
【分类号】：TP391.41

【相似文献】