Fixations in natural scenes: Interaction of image structure and image content

AbstractExplorative eye movements specifically target some parts of a scene while ignoring others. Here, we investigate how local image structure—defined by spatial frequency contrast—and informative image content—defined by higher order image statistics—are weighted for the selection of fixation points. We measured eye movements of macaque monkeys freely viewing a set of natural and manipulated images outside a particular task. To probe the effect of scene content, we locally introduced patches of pink noise into natural images, and to probe the interaction with image structure, we altered the contrast of the noise. We found that fixations specifically targeted the natural image parts and spared the uninformative noise patches. However, both increasing and decreasing the contrast of the noise attracted more fixations, and, in the extreme cases, compensated the effect of missing content. Introducing patches from another natural image led to similar results. In all paradigms tested, the interaction between scene structure and informative scene content was the same in any of the first six fixations on an image, demonstrating that the weighting of these factors is constant during viewing of an image. These results question theories, which suggest that initial fixations are driven by stimulus structure whereas later fixations are determined by informative scene content