基于语义增强的深度图像聚类方法研究  

RESEARCH ON THE DEEP IMAGE CLUSTERING METHOD BASED ON SEMANTIC ENHANCEMENT

作　　者：韩胜强 曲建华 Han Shengqiang;Qu Jianhua(Business School,Shandong Normal University,250358,Jinan,China)

机构地区：[1]山东师范大学商学院,济南250358

出　　处：《山东师范大学学报(自然科学版)》2024年第4期358-366,共9页Journal of Shandong Normal University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61876101,62102236)。

摘　　要：图像聚类旨在挖掘图像数据潜在的模式与规则,研究针对现有方法依赖内在特征而忽视外在语义特征致聚类效果欠佳的问题,提出新的深度图像聚类方法。该方法借助CLIP (Contrastive Language-Image Pretraining)模型挖掘语义特征,构建跨模态融合策略整合图像与文本信息,结合Kmeans算法构建深度聚类框架。在STL-10、CIFAR-10和CIFAR-20数据集上与15种已有方法及CLIP零样本分类方法对比实验,实验结果表明本文提出的图像聚类方法的聚类性能在多个指标上得到了显著提升。Image clustering aims to mine the potential patterns and rules of image data.Existing methods rely primarily on intrinsic features while neglecting external semantic features,resulting in an issue of suboptimal clustering.For this,a novel deep image clustering method is proposed here.This method excavates semantic features through CLIP model,constructs a cross-modal fusion strategy to integrate image and text information,and builds a deep clustering framework combined with Kmeans algorithm.Compared with 15 existing methods and CLIP zero sample classification method in STL-10,CIFAR-10 and CIFAR-20 data sets,the results show that the proposed method significantly improves clustering performance in clustering accuracy(ACC),normalized mutual information(NMI) and adjusting the Rand index(ARI) index.The proposed method provides a new avenue for image clustering and is expected to advance the development of related fields.

关 键 词：图像聚类 语义特征 多模态 深度聚类 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]