The visual computation of bounding contours
James H. Elder · eScholarship@McGill (McGill) · 1995
The problem of computing closed bounding contours from a visual image is addressed. The approach is motivated by the fact that many important image structures are heterogeneous and therefore cannot be segmented by traditional region-based methods based on homogeneity or smoothness constraints. Both psychophysical and computational aspects of the problem are studied. Psychophysical evidence is presented which shows that properties of bounding contour such as closure act as binding features for the segmentation of 2-D shapes, whereas texture (a region property) does not. These psychological results motivate an investigation of local and global issues in the computation of bounding contours for the purpose of figure/ground segmentation. Novel techniques for the local analysis of contour are presented which allow edges to be reliably detected and localized over a broad range of blur scale and contrast in images with shallow depth-of-field and shadows. Local estimates of contour blur are shown to be useful for computing depth segmentation from focal blur in complex images where the continuity assumptions required by existing Fourier techniques do not apply. It is hypothesized that in order to integrate contours fragmented by occlusion and insufficient contrast, the human visual system exploits a measure of geometric contour closure. The results of psychophysical experiments show that a measure for closure based on $L sb2$ norm of contour gaps in psychophysically consistent. The existence of this measure is evidence for the computation of geometric closure by the human visual system. These psychophysical findings motivate an algorithm for computing closed bounding contours as cycles of curve tangents. Tangent cycles extracted from a variety of real images are shown to correspond to the bounding contours of 2-D image structures, many of which are heterogeneous over their interior. Thus closure computations are seen to complement region-grouping methods by extending the class of structures which may be segmented to include heterogeneous image structures.