Improving Automatic Image Tagging Using Temporal Tag Co-occurrence

Existing automatic image annotation (AIA) systems that depend solely on low-level image features often produce poor results, particularly when annotating real-life collections. Tag co-occurrence has been shown to improve image annotation by identifying additional keywords associated with user-provided keywords. However, existing approaches have treated tag co-occurrence as a static measure over time, thereby ignoring the temporal trends of many tags. The temporal distribution of tags, however, caused by events, seasons and memes, etc, provides a strong source of evidence beyond keywords for AIA. In this paper we propose a temporal tag co-occurrence approach to improve AIA accuracy. By segmenting collection tags into multiple co-occurrence matrices, each covering an interval of time, we are able to give precedence to tags which not only co-occur each other, but also have temporal significance. We evaluate our approach on a real-life timestamped image collection from Flickr by performing experiments over a number of temporal interval sizes. Results show statistically significant improvements to annotation accuracy compared to a non-temporal co-occurrence baseline.