Using efficient visual exploration techniques to evaluate features for content-based image retrieval

We present a novel visual approach to evaluate, in a fast and effective way, the development of new image feature extraction techniques concerning content-based image retrieval. This new approach takes advantage of an interactive 3-dimensional visualization fed by the image features obtained through a given extraction technique under analysis. Using controlled test image datasets, the researcher can literally "see" the discrimination power of the image features. This new approach gives a very good insight of the behavior of a given image feature extractor algorithm, which are confirmed by the well-known precision and recall measurements. We applied the visual approach proposed onto a wavelet-based image retrieval system, which is supporting the development of a Picture Archiving and Communication System that allows one to retrieve images by content.