Robust texture features for still-image retrieval

A detailed evaluation of the use of texture features in a query-by-example approach to image retrieval is presented. Three radically different texture feature types motivated by i) statistical, ii) psychological and iii) signal processing points of view are used. The features were evaluated and tuned on retrieval tasks from the Corel collection and then evaluated and tested on the TRECVID 2003 and ImageCLEF 2004 collections. For the latter two the effects of combining texture features with a colour feature were studied. Texture features that perform particularly well are identified, demonstrating that they provide robust performance across a range of datasets.