Detecting shapes in noise: the role of contour-based and region-based representations