Content-based image retrieval via combination of similarity measures

A multiple (dis)similarity measure combination framework via normalization and weighting of measures is proposed to find suitable measure combinations in terms of retrieval accuracy and computational cost. In the combination of Manhattan and Hellinger distances, the computational time is more than 12 times faster and the retrieval accuracy improves or remains at the same level, when compared with Minkowski distance, a measure having the best retrieval accuracy in the single measure scenario. These performances are determined on a visual word based image retrieval system by using the Corel collections. Due to the reduction of computational cost and robustness of retrieval accuracy in this combination, applications include retrieval employing large number of images and categories in a database.