基于属性相关性的零样本学习  

Zero-shot learning based on attribute correlation network

作　　者：黄俊 刘孟奇 程泽凯 洪旭东 HUANG Jun;LIU Mengqi;CHENG Zekai;HONG Xudong(School of Computer Science and Technology,Anhui University of Technology,Maanshan,Anhui 243032,China;Institute of Artificial Intelligence,Hefei Comprehensive National Science Center,Hefei 230088,China)

机构地区：[1]安徽工业大学计算机科学与技术学院,安徽马鞍山243032 [2]合肥综合性国家科学中心人工智能研究院,合肥230088

出　　处：《中国科技论文》2023年第3期245-251,274,共8页China Sciencepaper

基　　金：国家自然科学基金资助项目(61806005);安徽高校协同创新项目(GXXT-2020-012,GXXT-2022-052);安徽高校科学研究重点项目(KJ2021A0373,KJ2019A0064)。

摘　　要：提出一种基于属性相关性的零样本学习(zero-shot learning,ZSL)方法。为了充分利用语义属性信息的内在关系,分别设计了基于均方误差(mean square error,MSE)的排名损失项和属性相关性损失项。排名损失项要求模型预测的视觉属性向量在所有类别属性向量中最靠近其真实类别,以学习到具有区分性的属性表示;属性相关性损失项使模型避免在可见类别样本上过拟合以提高模型在ZSL和广义零样本学习(generalized zero-shot learning,GZSL)任务上的泛化性能。在4个零样本学习基准数据集上进行了实验,验证了所提方法的有效性。A zero-shot learning(ZSL)method based on attribute correlation was proposed.In order to make full use of the intrinsic relationship of semantic attribute information,a ranking loss term based on mean square error(MSE)and an attribute relevance loss term were designed,respectively.The ranking loss term required that the visual attribute vector predicted by the model was closest to its true category among all category attribute vectors in order to learn a discriminative attribute representation.The attribute relevance loss term prevented the model from overfitting on the seen class samples and improved the generalization performance of the model on ZSL and generalized zero-shot learning(GZSL)tasks.Experiments were conducted on four zero-sample learning benchmark datasets and the effectiveness of the proposed approach was verified.

关 键 词：零样本学习 深度学习 图像分类 语义属性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP37[自动化与计算机技术—控制科学与工程]