Intermediate level picture interpretation using complete two-dimensional models

Abstract An intermediate level of computer vision is presented which is not general purpose as is the case with low level picture segmentation, nor is it based on semantic information processing as is usually prescribed for high level so-called cognitive processors. The input data to this stage is a partially segmented picture in that the given regions do not necessarily correspond to recognizable objects. In addition it is assumed that a complete two-dimensional picture model is available in a knowledge database containing information describing the objects appearing in the image, as well as their spatial structure. The picture interpretation involves two stages of processing, one local, and the other global. The first stage invokes local template matching whereby heuristic search is used to compare and match groups of adjacent regions against all stored object prototypes. This is followed by the second stage, an optimal search achieved by dynamic programming. Here the knowledge of the topological constraints is used to eliminate incorrect interpretations, resolving the competition between possible local object matches and the permissible global structure. The output of this intermediate level of processing is then an ordered list of possible symbolic interpretations for the input partially segmented regions. A confidence level is associated with each interpretation. Examples of face interpretation and object recognition are discussed.