The emphasis of this paper is low-level computer vision: the transformations from raw image input data to a segmentation into simple objects (or into lines, surfaces, colored patches, moving regions, etc.). The paper presents the beginnings of a theory of low-level computer vision. Quite a large number of ideas associated with this theory are discussed. Many of the individual ideas in the paper are not original, as the author states and carefully documents with 65 references. Many are new, as is the unification of the ideas into an approach to image segmentation.

Image segmentation has usually been viewed as grouping parts of intensity arrays. In this paper, segmentation is generalized to include grouping parts of intrinsic images (from an intrinsic image space) and parts of explicit representations of possible features and objects (from a feature space). Intrinsic images are calculated images of physical parameters; examples are images of velocity (optical flow), surface orientation, edges, and stereo disparity. The feature space of parameter values might be a space of colors or rigid body motion parameters. Possible grouping are represented as networks whose nodes signify parameter values. The relation between parts of an intrinsic image and the parameters in feature space are specified by a many to one constraint mapping. This approach to image segmentation potentially can handle noise and object occlusion.

A computational scheme to implement the theory is discussed. The scheme is based on massively-parallel cooperative computation between two groups of interconnected networks, and has particular virtue on a parallel non-von Neumann machine architecture.

Since this paper touches on so many important issues and ideas, it is recommended to the serious student or researcher working in the field of computer vision. For this reviewer, the discussions toward the end of the paper of ideas such as multiresolution relaxation methods, coupled computations, and spatial coherence were particularly illuminating.