Enriching Image Labels by Auto-Annotation: Simple but Effective and Efficient
Wei‐Lei Wang, Song Han, Ying Long Zhou, Yong Xing Jia, Yong Chen, Cao Qingxiang · 2012
Image search is popular and welcome currently, however the rareness of labels available may influence the performance of most commercial search engines built on traditional vector space model. To bridge the semantic gap between text labels and text queries, we propose one simple approach to extend existing labels using thesaurus. Different from naive approach where synonyms of tokens contained in a label are simply added, we employ three metrics to filter out those non-appropriate candidate labels (combination of synonyms): user log, semantic vector and context vector. The experiment results indicate that the proposed method has impressive performance, in term of effectiveness and efficiency.