Similitude-lnvariant Pattern Recognition Using Parallel Distributed Processing

AAAI 1987

TL;DR

This paper introduces a novel parallel distributed processing approach for recognizing objects in images, even when they are rotated, scaled, or partially occluded. It achieves this by using localizable features and a message-passing system to dynamically establish connections, allowing for fast, robust, and simultaneous recognition, localization, and segmentation of multiple objects.

What Problem Does This Paper Address?

Recognizing objects in images is a fundamental challenge in computer vision. Traditional methods often struggle with variations in viewpoint, scale, and occlusion, requiring explicit segmentation of objects before recognition. While connectionist models like interactive activation networks offer parallelism, they can be computationally expensive and often assume pre-segmented inputs. This paper tackles the problem of recognizing multiple, potentially overlapping or occluded objects in a single pass, without prior segmentation, and invariant to translation, rotation, and scale.

Methods

The core of the proposed method is a parallel distributed processing system that uses message passing instead of fixed connections. Key aspects include:

  • Localizable Features: The system relies on localizable and orientable features, specifically corners, as primitives. Each feature is associated with a processor.
  • Relational Coding: Models of objects are represented as a set of relative descriptions between features. For an object with N features, a model consists of 4 NxN arrays that specify the type, value, orientation, and direction of other features relative to a given feature's local coordinate frame. This relational coding inherently provides translation and rotation invariance.
  • Message Passing: Processors communicate by sending messages. When an image feature matches a model feature, it broadcasts messages to other processors, "expecting" other features of the same object to appear at specific locations and orientations.
  • Dynamic Connections: Connections are not fixed but are generated "on the fly" based on the messages received and the model information. This is described as programmable connections.
  • Winner-Takes-All: Each processor monitors support for feasible correspondences and selects the one with the highest support as the "correct" correspondence. This selection also determines the object's scale.
  • Segmentation: Segmentation is achieved through "support coincidence." If multiple correspondences predict the same set of feature locations, they are considered instances of the same object.

Datasets

Not discussed in the paper. The paper focuses on the algorithm's principles and demonstrates its capabilities with examples, including Figure 1 and Figure 3, which appear to be synthetic or illustrative rather than from a standard benchmark dataset.

Key Insights and Ideas

  • Simultaneous Recognition, Localization, and Segmentation: The most significant insight is that these three tasks can be performed concurrently and efficiently within a single parallel distributed processing framework, eliminating the need for a separate segmentation step.
  • Message Passing as a Flexible Communication Mechanism: Using message passing instead of fixed connections allows for dynamic and programmable interactions, making the system more versatile and efficient than traditional connectionist models.
  • Relational Coding for Invariance: Representing object features relative to each other, rather than in an absolute coordinate frame, is a powerful way to achieve scale, rotation, and translation invariance.
  • "Conspiracy of Partial Evidence": The system converges to a stable interpretation through mutual reinforcement of consistent feature correspondences, akin to a "conspiracy" where partial evidence from multiple features collectively drives a correct assignment.

Key Results

  • The algorithm is demonstrated to be fast, robust, and efficient.
  • Few iterations (typically no more than three) are required for recognition, with the first pass often yielding a correct assignment.
  • The system can recognize, localize, and segment multiple, superimposed, partially specified, or occluded objects.
  • The performance is primarily limited by the accuracy of input features, such as the reliability of image feature direction measurements.

Limitations and Future Work

  • Feature Vocabulary: Currently, the system primarily uses corners as features. The authors note that other localizable and orientable features can be used, and non-localizable/orientable features can serve as "confirmation" features.
  • Input Accuracy: The system's performance is sensitive to the accuracy of the input features, particularly their direction measurements.
  • Future Work: The paper suggests that this approach is well-suited for massively parallel architectures like the Connection Machine. Further exploration of different feature types and more complex object interactions could be future directions.

Summary updated

BibTeX
@inproceedings{1987similitude,
  title = {{Similitude-lnvariant Pattern Recognition Using Parallel Distributed Processing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure