Tagging Image by Merging Multiple Features in a Integrated Manner.

Xiaoming Zhang,Zhoujun Li,Wenhan Chao
DOI: https://doi.org/10.1007/s10844-011-0184-1
2011-01-01
Journal of Intelligent Information Systems
Abstract:Image tagging is a task that automatically assigns the query image with semantic keywords called tags, which significantly facilitates image search and organization. Since tags and image visual content are represented in different feature space, how to merge the multiple features by their correlation to tag the query image is an important problem. However, most of existing approaches merge the features by using a relatively simple mechanism rather than fully exploiting the correlations between different features. In this paper, we propose a new approach to fusing different features and their correlation simultaneously for image tagging. Specifically, we employ a Feature Correlation Graph to capture the correlations between different features in an integrated manner, which take features as nodes and their correlations as edges. Then, a revised probabilistic model based on Markov Random Field is used to describe the graph for evaluating tag’s relevance to query image. Based on that, we design an image tagging algorithm for large scale web image dataset. We evaluate our approach using two large real-life corpuses collected from Flickr, and the experimental results indicate the superiority of our proposed approach over state-of-the-art techniques.
What problem does this paper attempt to address?