基于FCM聚类的跨模态人物图像标注方法

摘要

With the explosive growth of multimodal people image data available, how to integrate multimodal information sources to achieve more accurate people image annotation becomes an important research issue. In this paper, a new framework is developed to support more precise automatic cross-modal people image annotation. It focuses on analyzing the associated text and image contents associated with multimodal people image and extracting the valuable information from both texts and images. For enhancing the whole performance of the cross-modal people image annotation approach, it particularly emphasizes on establishing an efficient measurement and optimization mechanism by Fuzzy C-Means Clustering Algorithm to verify the feasibility of matching between names and faces involved in multimodal people images. The experiments on a large number of official public data from Yahoo News have obtained very positive results.%提出一种新颖的基于模糊C均值（Fuzzy C-Means，FCM）聚类算法的跨模态人物图像标注方法，使用相关的人脸特征及文本语义，结合具有问题针对性的算法，建立有效的跨模态人物图像标注机制，进而对人物类图像标注性能进行提升。旨在于构建一种进行有效且准确的人物类图像标注的方法，进而在很大程度上提高人物类图像标注性能，在大规模带有相关联文本信息的人物类图像中，实现更加精确且有效的自动人脸-人名匹配。其贡献在于将人脸-人名匹配作为一种双模态媒体语义映射的问题进行处理，在双模态媒体（人脸图像与人名）的语义表达之间建立相应的关联分布，通过评估这种双模态媒体语义表达之间的相似关联性，进而针对人物图像标注最终衡量人脸与人名之间各种匹配方式的相对好坏。