Visual Concept Derivation from Natural Scenery Images Using Lexical Basis Functions, Multidimensional Scaling, and Density Clustering

Cognitive modeling of the visual concepts used by humans to classify images is a challenging task that first requires the characterization of images in terms of low-level or mid-level features that are salient to human visual perception. To complete the cognitive model, this characterization must then be correlated with higher-level concepts that are evoked in humans as they examine images. This paper presents a pilot experiment that involves characterization of outdoor scenery images in terms of mid-level visual content, as represented by words called lexical basis functions. To validate the list of lexical basis functions that are used for this purpose, the similarities between images (based on these basis functions) are correlated with subjectively perceived image similarities. To complete the cognitive model, multidimensional scaling and a novel multilevel density-based clustering algorithm are then used to cluster the images, and the resulting clusters are shown to correlate with salient high-level concepts that humans use to categorize images. Abstract. Cognitive modeling, visual concepts, visual percepts, lexical basis functions, multilevel density-based clustering, image indexing Cognitive modeling, visual concepts, visual percepts, lexical basis functions, multilevel density-based clustering, image indexing