Detection of Homogeneous Regions by Structural Analysis

Fumiaki Tomita, Masahiko Yachida, Saburo Tsuji

In the early 1970s, computer vision was primarily concerned with "blocks world" environments—scenes composed of simple, smooth geometric shapes. However, real-world scenes are rarely smooth; they are filled with texture. This paper presents a pioneering hierarchical method for image segmentation that moves beyond simple pixel intensity to analyze the underlying "structure" of a scene.

The Problem: Beyond Gray Levels

Traditional computer vision systems of the era relied on edge detection or gray-scale uniformity. These methods failed when faced with repetitive patterns. For instance, a patterned surface might appear as a uniform gray from a distance, but it is actually composed of distinct elements. The authors aimed to bridge this gap by creating a system capable of recognizing "homogeneous regions" based on the repetitive structure of their internal elements, rather than just their average brightness.

Methodology: Structural Analysis

The proposed system utilizes a multi-stage pipeline to decompose and then group image data:

  1. Atomic Regions: The image is first broken down into "atomic regions"—small, connected sets of pixels with nearly identical gray levels.
  2. Feature Extraction: Each atomic region is described by its shape (using moments $M_{20}, M_{02}, M_{21}, M_{12}, M_{11}$), size (area and perimeter), and density (distance to nearest neighbor).
  3. The Supervisor: An adaptive logic unit examines histograms of these descriptors. If a histogram shows distinct peaks separated by "deep valleys," that descriptor is selected as a candidate for partitioning the image.
Fig. 2: Illustration of how atomic elements (triangles) are grouped to define a homogeneous region.

By grouping these atomic regions, the system can define larger, structurally homogeneous areas. This process is iterative, allowing the system to find sub-structures within previously identified regions, a concept the authors call the "depth of interest."

Experimental Validation

The system was tested on custom-captured data using a vidicon TV camera. A primary test case involved a cube placed on a plane, where both the cube's faces and the background plane were covered in different repetitive patterns.

Input Scene 1: A patterned cube on a patterned plane used to test structural segmentation.

In this experiment, the system first used the perimeter descriptor to separate the cube from the background. It then performed a deeper partitioning on the cube region, using shape moments ($M_{11}$) to distinguish the three visible planes of the cube. The perspective distortion of the patterns on each face provided enough structural difference for the system to identify them as separate regions.

Historical Significance and Limitations

The core contribution of this work is the treatment of texture as a predictable arrangement of sub-patterns rather than statistical noise. By treating sub-patterns as "atomic regions," texture segmentation becomes a grouping problem. Furthermore, the "Supervisor" concept—where the system decides which feature is most useful based on the data—foreshadows modern adaptive machine learning techniques.

However, the method had notable limitations for its time:

  • Contrast Requirements: It required sharp outlines for atomic regions, struggling with blurry or low-contrast textures.
  • Computational Cost: Analyzing a 256x256 image took approximately 5 minutes on a HITAC 8350, a significant duration for the era.
  • Heuristic Simplicity: The Supervisor relied on simple histogram analysis, which the authors noted would need to become more sophisticated for complex real-world models.

Despite these hurdles, the paper laid the groundwork for structural texture analysis, moving the field toward a more nuanced understanding of visual scenes.

Summary updated

BibTeX
@inproceedings{tomita1973detection,
  author = {Fumiaki Tomita and Masahiko Yachida and Saburo Tsuji},
  title = {{Detection of Homogeneous Regions by Structural Analysis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1973},
}
Figure