基于语义对齐和重构的零样本学习算法认领

Zero-shot learning based on semantic alignment and reconstruction

在线阅读下载PDF

收藏分享导出

摘要为同时缓解零样本学习算法中固有的枢纽问题和域漂移问题,提出一种基于语义对齐和重构的零样本学习算法。以语义特征嵌入到图像空间的神经网络映射模型为基础,对模型添加语义原型和图像原型对齐的约束条件进一步缓解高维向量枢纽问题对标签预测的影响;对模型添加语义特征重构建的约束条件,缓解域漂移问题对识别正确率的影响。实验结果表明,所提算法在AwA和CUB数据集上达到了较优的识别正确率,验证了其有效性。 To alleviate the hubness problem and domain drift problem in the zero-shot learning,a reverse mapping method via semantic alignment and semantic reconstruction was proposed.Based on the neural network mapping model with semantic features embedded in image space,the constraints on semantic prototype and image prototype alignment were added to the model to improve the impact of high-dimensional vector hubness problems on label prediction.The constraint on semantic feature reconstruction was added to the model to improve the impact of domain drift problem on recognition accuracy.Experimental results show that the proposed algorithm achieves good recognition accuracies on AwA and CUB datasets,which verifies its effectiveness.

作者王紫沁杨维 WANG Zi-qin;YANG Wei(School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)

机构地区北京交通大学电子信息工程学院

出处《计算机工程与设计》北大核心 2021年第1期70-75,共6页 Computer Engineering and Design

关键词零样本学习语义对齐枢纽问题语义重构域漂移 zero-shot learning semantic alignment hubness problem semantic reconstruction domain drift

分类号 TP18 [自动化与计算机技术—控制理论与控制工程]