An effective strategy for querying image databases by color distribution

Content-based retrieval methods for image and video databases are receiving a great deal of attention as a result of the rapidly increasing availability of powerful devices which make it possible to acquire, process, store and transmit large collections of images and videos. Some research projects, such as QBIC (Query By Image Content) at IBM Almaden and CBVQS (Content-Based Visual Query System) at Columbia University, are exploring the use of low-level perceptual features, such as color, texture, shape and relative position, in indexing images [4,11]. These projects mainly address system efficiency, assuming that the databases to be queried are large, and that the user can then find the desired images by browsing a short list of some tens of candidates. In this paper we investigate the use of color for describing and retrieving images from a database. In fact, in several application fields-textile printing, ceramics and painting, for example-color information, to be of any use, must be faithfully described and rendered on the output device. If inappropriate color description and similarity evaluation schemes are adopted, many images that are actually similar in color may be missing from the short list. Moreover, since background and adjacent colors, image size, the observer's state of adaptation to lighting, etc., strongly influence our perception of color, target images can not be easily chosen among candidates by browsing through "thumbnail" representations of candidates. We present here an effective image retrieval strategy based on the fuzzy evaluation of color image similarity. In this method both the query and the database images are displayed in device-independent space with a limited palette of a perceptual significance. Image color distributions are represented by histograms, and a suitable similarity measure between histograms is also defined in order to model the perceptual similarity between their different colors (i.e. histogram bins). Experimental results on a database of some 120 images, which confirm the feasibility of the method, are also reported.