Human Perception-Driven, Similarity-Based Access to Image Databases

Similarity-based access to image databases assumes one or more similarity models. Although this assumption affects the retrieval precision of a system considerably, it is rarely described explicitly. Furthermore, because the similarity model is typically hard-coded into the system, it is very difficult if not impossible to use such a system for applications that do not fit the same similarity model. In this work, we develop a framework for designing similaritybased image access systems that are driven by human perception, and hence can be tailored for multiple, diverse applications. The driving components of the approach are Principal Components-based feature selection, perception modeling via psychophysical experiments and Genetic Algorithm-driven distance function optimization. While our framework is general and flexible, we demonstrate the application in a particular image access scenario: Shapebased retrieval of skin lesion images. The experimental results show that, by incorporating human perception of similarity into the system, retrieval performance may be significantly improved.