A Dynamic Pattern Recognition Method Using the Preview Controlled Saccadic Movement of the Detector

Masahiro Mori, Kunikatsu Takase, Harukazu NishioIJCAI 1969

Abstract: This classic research from the University of Tokyo (circa late 1960s) presents a bio-inspired approach to computer vision that mimics the "saccadic" movements of the human eye. By controlling the speed of a mechanical detector based on the curvature of the lines it tracks, the system achieves recognition that is naturally invariant to an object's size, rotation, and slight deformations. This paper explores the transition from static pixel-based templates to dynamic, time-normalized feature extraction.

Introduction

In the early era of computer vision, recognizing handwritten shapes was a significant challenge due to variations in position, scale, and orientation. Traditional scanning techniques relied on fixed coordinate systems, making them fragile to even minor shifts. Mori, Takase, and Nishio argued that biological systems avoid these pitfalls because they do not perceive the world statically. Citing experiments on human ocular movement and the visual systems of amphibians, the authors proposed that motion is a fundamental requirement for sight. Their goal was to build a mechanical system that uses motion to extract the topological "essence" of a shape.

The Preview Controlled Tracking System

The researchers developed a "Preview Controlled" tracking system. Instead of a static camera, they utilized a mobile optical detector ($P_s$) equipped with two "preview" photocells ($P_{pi}$ for inside and $P_{po}$ for outside). These preview cells rotate rapidly as the detector moves along a line, generating a sequence of pulses whenever they cross the pattern's edge.

Schematic of the preview controlled detector mechanism showing the main sensor $P_s$ and preview photocells $P_{pi}$ and $P_{po}$ over a pattern segment.

Curvature-Dependent Speed and Normalization

The core innovation of this system is the mathematical relationship between tracking speed ($V$) and the radius of curvature ($R$):

$$V = \frac{K_1 R}{1 + K_2 R}$$

This control law forces the detector to slow down for tight corners and speed up on straightaways, much like a human driver. This mechanism effectively turns "size" into "time." Because the detector moves faster on the larger, flatter curves of a big shape and slower on the tighter curves of a small shape, the total tracking time remains roughly constant for similar topologies. This achieves scale invariance without complex coordinate transformations.

Feature Extraction and Recognition

The pulse sequences from the rotating preview cells are fed into a logic circuit based on a tree algorithm. Specific sequences of "Inside" (I) and "Outside" (O) pulses correspond to fundamental topological features such as straight lines, arcs, angles, intersections, or endpoints.

A pattern is then recognized as a time-ordered sequence of these features. For example, a "0" is identified as a single continuous arc, while a "4" is recognized as a specific sequence of straight lines and intersections. This feature-based logic makes the system highly robust to the slight deformations common in human handwriting.

Experimental Results

The authors evaluated the method using handwritten numerals (0-9). The results demonstrated that the system could successfully identify numerals regardless of their position on the page or their orientation.

Comparison of tracking signatures for two different handwritten '2's, showing how different physical inputs result in similar velocity-time profiles.

As shown in the comparative analysis of handwritten "2"s, the "Tracking Time vs. Speed" (T-V) graphs remain remarkably consistent across different samples. Even when one numeral is larger or more slanted than another, the temporal "signature" of the movement remains recognizable, proving the effectiveness of the normalization technique.

Conclusion and Legacy

While the mechanical hardware of the late 1960s limited the speed of the physical detector, the conceptual framework remains highly relevant. The authors correctly identified that saccadic movement could serve as a powerful "front-end" for pattern recognition, providing normalized data to more complex neural networks. This work serves as a foundational reminder that bio-inspired motion can solve complex invariance problems that are often computationally expensive in static image processing.

Summary updated

BibTeX
@inproceedings{mori1969dynamic,
  author = {Masahiro Mori and Kunikatsu Takase and Harukazu Nishio},
  title = {{A Dynamic Pattern Recognition Method Using the Preview Controlled Saccadic Movement of the Detector}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1969},
}
Figure