Edge Finding, Segmentation Of Edges And Recognition Of Complex Objects

Yoshiaki Shirai

Abstract

In the early 1970s, computer vision systems were largely confined to the "blocks world," successfully identifying simple polyhedra but struggling with the complexities of curved, real-world objects. This paper presents a pioneering hierarchical system that bridges this gap. By combining a flexible edge-tracking algorithm with a high-level "supervisor" that iteratively adjusts detection thresholds based on object knowledge, the system demonstrates the ability to recognize complex items like telephones and lamps in cluttered office environments.

Introduction

The primary challenge addressed in this work is the transition from recognizing flat-faced objects (cubes, pyramids) to curved, real-world objects. Traditional systems of the era were strictly serial: they extracted features first and then attempted recognition. This approach was brittle; if initial edge detection failed due to noise or poor lighting, the entire pipeline collapsed.

The author identifies two critical bottlenecks: the mathematical representation of non-linear curves and the lack of feedback between high-level object knowledge and low-level sensing. To solve these, the paper proposes a system where recognition is not just the final step, but an active participant in the sensing process.

Methodology

The system architecture is divided into four interacting modules:

Edge Finding

The system employs a dual-operator approach for edge detection. A 2D operator first calculates gradient directions for unknown edges. Once a direction is established, a faster 1D operator tracks the edge. Edge points are classified into three categories: Type B (plane boundaries) and Type L/R (left or right sides of curved surfaces).

The edge detection operator uses different neighborhood modes to calculate gradients and track edges.

Segmentation and Recognition

Rather than treating edges as isolated pixels, the system groups them into meaningful segments. It identifies "knots"—points of abrupt curvature change—and fits the intervening segments into straight lines or elliptic curves using Deming’s least-squares method.

For recognition, the system uses two distinct strategies:

  1. Graph Matching: For simpler objects, it searches for common subgraphs between the observed scene and stored models.
  2. Procedural Approach: For complex objects like telephones, it follows a specific search procedure—for instance, identifying the circular dial first to serve as a "seed" for locating the rest of the handset and base.

The Supervisor

The "Supervisor" acts as the system's control center. It initially directs the edge finder to look for the most reliable, high-contrast edges. If these initial features are insufficient for recognition, the supervisor instructs the edge finder to lower its reliability thresholds and re-examine specific regions of interest for "fainter" edges that might complete an object hypothesis.

Key Results and Discussion

The system was evaluated using real-world office scenes captured at a resolution of 256x256 pixels with 6-bit light intensity. Despite the limited hardware of the time, the system successfully identified lamps, book stands, and telephones in cluttered scenes.

A comparison between the raw input data (left) and the system's extracted edge segments for a telephone (right).

Processing a scene with three objects took approximately 4 minutes. Notably, the "supervisor" approach proved to be twice as efficient as a "blind" search, demonstrating that top-down guidance significantly reduces the computational burden of low-level feature extraction.

Conclusion and Future Work

This work highlights that vision is most effective as a two-way street between sensing and recognition. While the system faced limitations—such as sensitivity to object rotation in graph matching and the lack of a fully implemented "zoom" feedback loop—it laid the groundwork for modern hierarchical vision architectures. Future improvements suggested by the author include more robust model libraries and enhanced feedback mechanisms to handle even greater scene complexity.

Summary updated

BibTeX
@inproceedings{shirai1975edge,
  author = {Yoshiaki Shirai},
  title = {{Edge Finding, Segmentation Of Edges And Recognition Of Complex Objects}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure