Putting Gestalt Laws at Work

The inherent complexity and ambiguity of scene properties renders vision extremely difficult without taking advantage of the statistical regularities of natural images. The most basic forms of these image regularities are captured by the Gestalt laws of perceptual organization [5, 15, 21]. This paper reports two versions of a multiscale and multiorientation artificial recognition system in which some of these Gestalt laws: good continuation, cocurvilinearity, proximity, and constancy of curvature (smoothness) are explicitly implemented. The recognition performance of these two ‘Gestalt enhanced’ models, of humans and of a baseline model where no Gestalt principles were implemented are compared on several sets of line-drawn images. It was found that the performance of the two ‘Gestalt enhanced’ recognition systems not only increased significantly compared with the baseline model, but their performance also became qualitatively much closer to that of humans. There was no significant difference found between the performance of the two enhanced models with slightly different Gestalt law implementations.