COMPARATIVE ANALYSIS OF POPULAR FACIAL RECOGNITION TECHNIQUES IN THE IMAGES

Dmitry Kuznetsov, V.a. Damm, Andrey Kuznetsov, R.b. TregubovResearch 2019 · DOI

Abstract

Facial recognition systems are increasingly deployed in high-security environments, yet they often struggle with "unstable shooting conditions" such as poor lighting, camera noise, and the natural aging of subjects. This paper provides a comparative review of five prominent facial recognition algorithms: Elastic Bunch Graph Matching, Principal Component Analysis (PCA), the Viola-Jones method, Support Vector Machines (SVM), and the novel Face Image Barcode method. By synthesizing results from 22 distinct studies, we evaluate these methods against real-world stressors. Our findings highlight the superior performance of the "biological barcode" method in handling facial aging, while the Viola-Jones method remains the most robust against noise and head rotation.

Introduction

In the realm of multimodal authentication, facial recognition serves as a cornerstone. However, the transition from controlled laboratory settings to real-world applications introduces significant challenges. Systems must operate under variable brightness, handle sensor noise, and account for head tilts. Perhaps most critically, they must remain effective as a subject ages over decades—a factor that typically necessitates frequent and costly database updates. This paper aims to identify the most robust algorithmic approaches for high-security "informatization objects," where reliability and administrative efficiency are paramount.

Methodology

The standard facial recognition process follows a linear pipeline consisting of four primary stages: detection, alignment, feature extraction, and comparison.

Standard facial recognition pipeline: detection, alignment, feature extraction, and comparison.

Within this framework, five specific techniques are analyzed:

  1. Elastic Bunch Graph Matching (EBGM): Represents faces as deformable graphs using Gabor filters for feature extraction.
  2. Principal Component Analysis (PCA): A statistical method that reduces image dimensionality into "eigenfaces."
  3. Viola-Jones Method: A rapid detection technique using Haar-like wavelets and cascaded classifiers.
  4. Support Vector Machines (SVM): A machine learning approach that classifies faces by finding optimal hyperplanes in high-dimensional space.
  5. Face Image Barcode Method: A structural approach that extracts brightness gradients from horizontal facial lines to create a "biological barcode" (standardized as EAN-8).

Comparative Performance Analysis

The effectiveness of these algorithms diverges sharply when subjected to environmental and temporal variables. While PCA is computationally elegant, it proves to be the most fragile under non-ideal conditions. In contrast, structural methods like graph matching and the barcode method handle facial dynamics and expressions with greater stability.

The Impact of Aging

A significant finding of this meta-analysis is the "aging advantage" of the barcode method. Most systems require periodic re-enrollment as facial features change over time. However, simulations modeling aging from age 20 to 60 showed that the barcode method maintained 100% effectiveness. By focusing on structural invariants—such as the relative positions and gradients of eyebrows, eyes, and lips—this method captures features that remain stable across decades.

Robustness to Noise and Rotation

While the barcode method excels in temporal stability, the Viola-Jones method remains superior for handling spatial instability. In scenarios involving significant camera noise or head rotation (XY plane), Viola-Jones achieved 94% accuracy, outperforming the structural and statistical alternatives.

Results Summary

The following table consolidates performance metrics across four critical stress factors:

Condition Barcode Method Graph Matching PCA Viola-Jones SVM
Facial Dynamics 100% 90% 75% 98% 97%
Brightness Change (40%) 50% N/A 50% 70% 39%
Noise & Rotation (XY) 78% 81% 86% 94% 80%
Aging 100% N/A 50% N/A N/A

Conclusion

The choice of a facial recognition algorithm depends heavily on the specific operational environment. For military or federal security infrastructure, the "biological barcode" method offers a significant advantage by minimizing the administrative overhead of re-training classifiers due to aging. However, for systems prone to high sensor noise or uncooperative subjects (rotation), the Viola-Jones method remains the most reliable. Future developments in robust authentication likely lie in hybrid systems that combine the spatial detection strengths of Viola-Jones with the long-term structural stability of the barcode method.

Summary updated

Abstract

. . . , , -, , - .
BibTeX
@inproceedings{kuznetsov2019comparative,
  author = {Dmitry Kuznetsov and V.A. Damm and Andrey Kuznetsov and R.B. Tregubov and R.B. Tregubov},
  title = {{COMPARATIVE ANALYSIS OF POPULAR FACIAL RECOGNITION TECHNIQUES IN THE IMAGES}},
  booktitle = {Research},
  year = {2019},
  doi = {10.18413/2518-1092-2019-4-4-0-1},
}
Figure