A Computational Model for Face Location Based on Cognitive Principles

Venu Govindaraju, Sargur N. Srihari, David B. SherAAAI 1992

Introduction

In 1992, researchers at the University at Buffalo proposed a novel bio-inspired approach to face localization, a foundational challenge in computer vision. While contemporary research often focused on face identification—matching a pre-cropped face to a database—this paper addressed the more difficult task of locating faces within unconstrained, cluttered environments like newspaper photographs. The authors posited that the human face is a "special" object requiring a dedicated computational model rather than general-purpose object recognition techniques.

Methodology: Hypothesis Generate and Test

The core of the system is a "Hypothesis Generate and Test" paradigm structured around cognitive principles derived from infant visual development.

Hypothesis Generation: The "Infant" Approach

Research in child psychology suggests that one-month-old infants prioritize head outlines and hairlines before noticing internal features. Mimicking this, the system first performs edge detection using the Marr-Hildreth operator to identify "feature curves" representing the top and sides of a head.

The multi-stage process of face localization, from edge detection to final hypothesis verification.

Feature Extraction and Scaling

To handle varying face sizes, the model employs anthropometric data (specifically the "Golden Ratio") to define ideal face proportions. This allows the detected curves to scale mathematically. The system focuses on edges rather than raw pixels to capture essential shape information while remaining robust against the noise typical of newspaper imagery.

Edge-based feature curves used to hypothesize potential face locations.

Hypothesis Testing and Caption Integration

Once potential locations are flagged, the system mimics a two-month-old's development by searching for finer details, such as pairs of dark regions corresponding to eyes. This verification step is constrained by information extracted from the photograph's caption via a system called "Piction." The caption provides a "target count" of individuals and their approximate spatial arrangement, allowing the system to prune false positives that do not fit the narrative context.

Experimental Results

The model was evaluated using a dataset of 60 newspaper photographs from The Buffalo News (1991), featuring real-life clutter and varied lighting.

  • Overall Success: The system successfully located every face mentioned in the caption in 73% of the test cases.
  • Detection Rate: Out of 90 total faces, the system correctly identified 68 (76%).
  • Single-Face Accuracy: In images containing only one person, accuracy rose to 83%.

By restricting the search for internal features to small hypothesized regions, the system achieved significant computational efficiency compared to exhaustive search methods.

Conclusion

The study demonstrates the effectiveness of incorporating chronological perceptual sequences into computer vision. By prioritizing head shapes before internal features and leveraging multi-modal cues from captions, the researchers developed a robust framework for face localization in unconstrained scenes. While limited by pose constraints and sensitivity to extremely busy backgrounds, this work laid important groundwork for specialized object detection systems.

Summary updated

BibTeX
@inproceedings{govindaraju1992computational,
  author = {Venu Govindaraju and Sargur N. Srihari and David B. Sher},
  title = {{A Computational Model for Face Location Based on Cognitive Principles}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}

Loading related content...

Figure