基于动态图卷积的图像情感分布预测  

Dynamic graph convolutional neural network for image sentiment distribution prediction

作　　者：苏育挺 王骥 赵玮 井佩光 SU Yu-ting;WANG Ji;ZHAO Wei;JING Pei-guang(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China;Tianjin International Engineering Institute,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072 [2]天津大学国际工程师学院,天津300072

出　　处：《吉林大学学报（工学版）》2023年第9期2601-2610,共10页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(61802277)。

摘　　要：针对图像情感分布学习中,视觉特征与高阶情感语义之间存在语义鸿沟以及情感标签具有主观性和模糊性的问题,提出了一种情感语义动态图卷积网络模型。该模型通过情感激活模块自动定位情感语义区域,从而有效挖掘契合情感语义的内容表征;通过动态图卷积模块自适应地捕获图像情感标签之间的语义关联性;最终构建并行结构输出联合局部语义和标签相关性的情感预测分布。在3个公开情感数据集上的实验结果证明了本文算法在图像情感分布预测任务中的有效性。Aiming at the problem that there exists semantic gap between visual features and high-level emotional semantics and the subjectivity and ambiguity of emotional labels in image sentiment distribution learning,this paper proposes an Emotional Semantic Dynamic Graph Convolution Network(ESDGCN).In this framework,the Emotion Activation Module(EAM)is constructed to automatically locate the emotional semantic regions to effectively mine the content representation that fits the emotional semantics.In addition,the Semantic Dynamic Graph Convolution Network(SDGCN)is to adaptively capture the semantic relevance between labels.Finally,we adopt the parallel structure to jointly consider local semantic emotional information and label correlations.Experimental results on three open emotional datasets demonstrate the effectiveness of the proposed method.

关 键 词：信息处理技术 视觉情感计算 动态图卷积 标签分布式学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]