Junk Image Filtering via Multimodal Clustering for Tag-based Social Image Search

Tag-based social image search is now very popular for accessing large numbers of collaboratively tagged images on social image sharing platforms. Unfortunately, because of the uncertainty of the relatedness between images and their annotation tags, a very large number of junk images are returned by such search engines without filtering. In this paper, to filter out junk images, the result images for a unique tag query are classified into the relevant category and junk category. Multimodal clustering is automatically performed to construct these two clusterings by leveraging weakly tagged social images. Our experiments on public Flickr social images have shown that multimodal clustering for social images can filter out junk images more effectively.