Color-induced image representation and retrieval

摘要

In this paper, we describe a framework for pictorial content representation, query formulation and image retrieval based on color distributions. Image content is described through a set of color histograms, each relative to a region with a homogeneous color distribution. The representation also includes geometric features induced by the color distribution based segmentation process. A metric of similarity between images using the above representation and generalizing the histogram intersection operator is defined, allowing to perform retrieval by color distribution content. User query images are interactively created and modified through a graphic environment featuring color sketching, image examples and relevance feedback. The user is also provided with two novel query composition and refinement tools, based respectively on an internal query memory and on a direct manipulation of internal query representation. Examples of operation are provided, and experimental results are reported indicating that, thanks to the superiority of multiple distributions over global color histogramming, the quality of image retrieval is in good accordance with human expectation.