Natural scene categorization by global scene properties: Evidence from patterns of fMRI activity

Abstract

Human scene categorization is remarkably rapid and accurate, but little is known about the neural representation mediating this feat. While previous studies on neural representation of scenes have focused on basic level scene categories, here we examined whether the neural representation of scenes reflect global properties of scene structure, such as openness of a space, or properties of surfaces and contents within a space, such as naturalness. In an fMRI study, human participants performed a one-back task on blocks of images of four scene groups: Open Natural images, Closed Natural images, Open Urban images, Closed Urban images. Each image group included multiple basic level categories. For example, Open Natural images included open views of fields, oceans and deserts; while Open Urban images included open views of highways, parking lots, and airports. For each participant, we defined regions of interest (ROIs) of the parahippocampal place area (PPA), the fusiform face area (FFA), lateral occipital complex (LOC) and V1. Multivariate pattern analysis was applied to voxels within each ROIs, and split-half pattern correlation and Euclidian distances across voxel activations were calculated (Haxby et al., 2001). We observed high identification accuracy in the PPA and V1, but not in the FFA and LOC. Most interestingly, when the correct identification failed in the PPA, the confusion was between images with the same layout rather than between images with the same content. For example, Open Natural images were often highly correlated with Open Urban images, but rarely with Closed Natural images. These results suggest that a critical component of scene representation in the brain is the coding of global properties of spatial layout.