Edge co-occurrences are sufficient to categorize natural versus animal images

Determining the category of a visual scene, such as whether it contains an animal, has been proposed to use a hierarchy of brain areas that progressively analyze the scene at increasing levels of abstraction, from contour extraction to low-level object recognition and finally to object categorization (Serre et al. 2007, PNAS 104:6424-9). We explore an alternative hypothesis that the a V1-based representation of the statistics of edge co-occurences are sufficient to perform a rough yet robust (translation, scale, and rotation invariant) scene categorization. Using a scale-space analysis coupled with a sparse coding algorithm, we first achieved detailed and robust extraction of edges in three datasets of natural images in different categories (natural, man-made, or containing an animal). We then computed the “association field” for each dataset by computing the statistics of edge co-occurrences, as in Geisler et al. (Vision Res. 2001). These statistics differed strongly between datasets, with images of man-made objects having more straight and parallel configurations, and images of animals having more co-circular (curved) configurations. We show that this geometry is sufficient to distinguish the image category – given only this geometry, a simple classifier gives performance similar to that of humans in rapid-categorization tasks. These results suggest new algorithms for image classification that exploit correlations between local image structure and the underlying semantic category, and they challenge widely held assumptions about the flow of computations in the visual system.