Texture, human perception, and information retrieval measures

Content-based image retrieval (CBIR) uses image properties such as colour, texture and shape. Although colour has been used successfully, texture can be even more significant, both for human perception and for CBIR. We present the results from a human study with 30 volunteers that ranks the “most like” images for each of the Brodatz textures. In all cases, only a limited number of the other textures were considered to be in any way similar. This perceptually-derived ranking is then used as a basis for relevance judgements, and retrieval performance measures calculated for the ranking given by each of ten computational texture-based methods. Precision, recall and Faloutsos’ AVRR measure are calculated, and the averages plotted for each method. These plots provide some guidance on the optimum number of retrievals. However, it is also necessary to consider the relative ordering of the retrievals, and the nature of the nonrelevant images retrieved in each case, when considering system usability.