Learning Label Set Relevance for Search Based Image Annotation

As a way to facilitate image categorization and retrieval, automatic image annotation has received much research attention. Traditional web image annotation methods often estimate the label relevance to image by the most common labels´ frequency derived from its nearest neighbors, and neglect the relevance of the assigned label set as a whole. We propose in this paper a novel search based image annotation method by learning label set relevance, which aims at annotating large scale image collections in real environment. "Label set"-to-image relevance and label set correlation are formulated into a joint framework. Measures that can estimate the label set relevance are designed. The assigned label set provide a more precise description of the image´s content. To reduce the complexity, a heuristic algorithm is introduced to annotate image accurately and efficiently in large scale web image set. Experiments on real world web dataset demonstrate the general applicability of our algorithm in web image annotation. In comparison to state-of-the-art, the proposed method achieves excellent performance.