The ultimate goal of computer vision is image understanding, in other words, knowing what is within an image at every x and y point. A complete computer vision system should be able to segment the image into homogeneous portions, extract regions from the segments that are single objects, and finally output a response as to the locations of these objects and what they are.

The framework for image understanding consists of three not necessarily separate processes. A representative computer vision system is shown in figure 1. In the first process of figure 1, image segmentation is performed. Image segmentation consists of dividing the image into homogeneous portions that are similar based on a correlation criterion. In this document, image segmentation will also be known as the low-level vision (LLV) process. In figure 1, the second or intermediate-level vision (ILV) process performs region extraction. Region extraction may receive the results obtained during an LLV process or the original image itself. With this information, the ILV process attempts to represent image objects from hypothesized objects. Subsequently, the third process of figure 1 performs image understanding operations based on the extracted regions provided as input. The hypothesized image understanding operation will be known as the high-level vision (HLV) process.

Most computer vision research for approximately the past 30 years has focused on LLV processes. Only recently has some attention been devoted to furthering the knowledge of ILV processes, primarily using LLV methods. These LLV techniques work well if the image properties are uniform or homogeneous (e.g., same gray level, texture). However, these methods are inapplicable for regions whose image properties are nonuniform or heterogeneous. Therefore, what is needed is a new technique specific to the ILV goal that will extract regions of nonuniform image properties.

Accordingly, the main focus of this research was on the ILV process. Hence, an energy minimization technique is provided that recognizes compact-closed objects represented in polar coordinate form. These compact-closed objects are used to characterize a Markov Random Field (MRF), which is incorporated into an energy minimization function. An initial high-level hypothesis is provided by a simulated HLV process (i.e., image analyst or human). A combinatorial optimization technique, known as tabu search, then provides the means for driving the energy function to its minimum state. This research also showed how the minimum energy state corresponds to an MRF state of highest probability (i.e., Gibbs Distribution).

A smaller set of results is provided showing the algorithm’s capability to extract a quadrilateral region from an image. In one case, the quadrilateral region is a synthetic representation of a building. In the other two cases, the quadrilateral region is a real-world object in the form of a building. Figure 2 is the synthetic case used for experimentation. Figures 5 and 7 are the real-world cases used for testing.