空天预警研究学报2025，Vol.39Issue(6)：431-435,5.DOI:10.3969/j.issn.2097-180X.2025.06.008
基于生成对抗网络的预警装备文本-图像联合抽取方法
A GAN-based joint extraction method for text-image of early warning equipment
冯露君 1方其庆 2胡亚慧 2冯雪峰2
作者信息
- 1. 空军预警学院,武汉 430019||31811部队,重庆 400041
- 2. 空军预警学院,武汉 430019
- 折叠
摘要
Abstract
Aiming at the problems of separated text-image modalities,weak correlation between entities and visual features in early warning equipment maintenance support,as well as strong annotation dependence and poor generalization ability of the traditional methods,this paper proposes a joint extraction method based on gener-ative adversarial network(GAN).Firstly,a"generator-discriminator-feature fusion"architecture is constructed.Then,the text semantics and image visual features are extracted through BERT and ResNet50,and unified to a 512-dimensional space through feature fusion.Finally,the generator is based on text features and random noise to generate visual candidates;the discriminator verifies the"entity-visual"matching degree through attention fusion and classification;the two modules of generator and discriminator carry out adversarial optimization.Experimen-tal results show that the F1-score of text entity extraction of the proposed method reaches 93.2%,which is 4.17%higher than that of the BERT-BiLSTM-CRF method,that the visual matching accuracy rate is 92.3%,which is 18.6%higher than that of the traditional cross-modal matching methods,and that the F1-score of unlabeled data is 89.8%.关键词
预警装备/生成对抗网络/多模态抽取/文本-图像联合抽取/知识图谱Key words
early warning equipment/generative adversarial network(GAN)/multimodal extraction/text-im-age joint modeling/knowledge graph分类
信息技术与安全科学引用本文复制引用
冯露君,方其庆,胡亚慧,冯雪峰..基于生成对抗网络的预警装备文本-图像联合抽取方法[J].空天预警研究学报,2025,39(6):431-435,5.
