Abstract: This paper presents an efficient method for facial recognition by integrating fuzzy theory and Shannon entropy. This combination simplifies the recognition process compared to traditional methods. Shannon entropy is utilized to calculate the ratios of facial elements, ensuring scale invariance, while fuzzy theory determines the membership degree of these entropy values. The proposed system demonstrates high learning performance, requiring only two data points per individual, and focuses on stable facial features to maintain accuracy over time.
Introduction
Modern facial recognition often relies on computationally intensive models like Hidden Markov Models (HMM) or Eigenfaces. While effective, these systems frequently struggle with high complexity and sensitivity to environmental conditions, requiring large datasets for training. In "Fuzzy and Entropy Facial Recognition," Jae Jun Lee and Taeseon Yoon from the Hankuk Academy of Foreign Studies propose a lightweight alternative. Their approach leverages the mathematical rigor of Shannon entropy and the flexibility of fuzzy logic to create a system that is scale-invariant and requires minimal training data.
Addressing the Complexity of Traditional Recognition
Traditional "crisp" logic (binary true/false) often fails to capture the subtle nuances of human faces, which vary due to lighting, expression, or aging. The authors sought to eliminate these complications by focusing on stable facial ratios and "fuzzy" membership degrees. This allows for a system that is simple to implement yet robust against changes in image scale.
The Proposed Methodology
The recognition pipeline consists of several key stages:
1. Feature Extraction
The system identifies geometric distances between key landmarks, such as the distance between eyes, nose-to-mouth length, ear-to-ear width, and eyebrow length.
2. Shannon Entropy Calculation
Instead of processing raw pixel data, the system applies Shannon entropy to the extracted feature sets. Entropy measures the unpredictability or complexity of data.
By using entropy, which deals with ratios, the system becomes naturally scale-invariant. The ratio of eye distance to nose length remains constant regardless of whether the face appears large or small in the frame.
3. Fuzzy Membership
The calculated entropy is processed through a bell-curve fuzzy membership function. This function determines the "degree of membership" (a value between 0 and 1) of an input face relative to a stored template.
This allows the system to quantify how closely a face matches a template, rather than providing a binary yes/no answer, accommodating natural variations.
4. Face Line Subtraction and Hybrid Scoring
A secondary check involves drawing a face line (excluding hair to avoid noise) and subtracting the area of a new input from a stored template to calculate a similarity factor. The final similarity score is a weighted combination of the fuzzy entropy score and this area subtraction result.
Key Insights
- Scale Independence: By focusing on ratios via Shannon entropy, the system avoids the common pitfall of resolution or distance sensitivity.
- Minimalist Learning: The algorithm is remarkably efficient, establishing initial variables with just two images per person.
- Iterative Refinement: The system uses a "Constant K" that updates as it processes more images, allowing it to adapt to a user's face over time, similar to modern mobile security systems.
Results and Future Directions
The authors report that the algorithm is designed to identify matches within a 95% to 100% similarity range. Its simplicity leads to superior computational efficiency compared to data-heavy deep learning models.
However, the current version has limitations. It is primarily designed for front-facing images and cannot yet handle significant head tilts or rotations. Future work will explore incorporating 3D components into the 2D recognition process to further enhance accuracy and robustness against environmental factors.
Summary updated