Integrating Multiple Types of Features for Event Identification in Social Images
Xiaoming Zhang,Zhoujun Li,Xueqiang Lv,Xiaoming Chen
DOI: https://doi.org/10.1007/s11042-014-2436-x
IF: 2.577
2015-01-01
Multimedia Tools and Applications
Abstract:With the rapidly increasing popularity of social media sites, a large amount of user-generated data has been injected into the web. The data include a wide variety of real-world events. As a consequence, especially for social multimedia objects, it has become increasingly difficult to allow the browsing and organization of multimedia collections in a more effective manner. The approach we propose in this study addresses this problem, thus enabling the browsing and organization of multimedia collections in a natural way, i.e., by events. There have been some research studies on this problem. However, most of the previous approaches merge multiple types of features (e.g., textual content, visual content, user information and temporal information) of social media using a relatively simple mechanism. In this study, we merge multiple types of features in an integrated manner to identify the event associated with user-contributed social multimedia objects. We exploit the correlations between different types of features, i.e., textual content, visual content, user information and temporal information, to classify new social multimedia objects into their corresponding event categories. We accomplish this through a feature correlation graph ( FCG ) that uses features as nodes and the correlations among these features as edges for each event and individual multimedia object. We then employ a probabilistic model based on Markov random field to connect each new multimedia object with the correct event. We evaluate the algorithm on large-scale, real-world datasets of event images downloaded from Flickr, and the experimental results confirm the superiority of our approach over state-of-the-art approaches.