A semantic relatedness-based solution for reducing missing problem in TBIR

In text-based image retrieval, matching is a technique that retrieves for a concept query Q, images annotated with Q. The result performance is very influenced by the annotation quality. Since it is difficult to have a well-annotated data set, the retrieval neglects many relevant images simply because they are not annotated with the query concept (i.e. missing problem). In this paper, we propose a solution that considerably minimises such a problem, by integrating the semantic relatedness between concepts into the retrieval. We compute the semantic relatedness between pairs of concepts from Wikipedia articles. We use term frequency - inverse collection term frequency weighting scheme and the cosine similarity. After evaluating the obtained values, using the human judgement benchmark WordSimilarity-353, we incorporated them into image retrieval task. The experimental results on Corel 5K data set clearly show the ability of the proposed method in detecting missing images, compared with matching and some literature works.