Multi-scale Block Weber Local Descriptor for Face Recognition

Weilan Pu, Xiaohong Zhang2015 · DOI

Abstract

In this paper, we provide a novel local descriptor Multi-scale Block Weber Local Descriptor (MBWLD) for face recognition. It is a LBP-based descriptor which combines MBLBP with WLD for integrating both their advantages. Firstly, Multi-scale block filters is used on the face image to calculate the average gray-values of sub-regions. Secondly, differential excitation and orientation of WLD is extracted using the average neighbors and center. MBWLD extracts both the micro and macro patterns from face image, and becomes more descriptive and robust to face representation. The feasibility and availability are validated by comparing with the standard LBP, MBLBP and WLD on FERET database in expression, illumination and age changes. The experimental result shows that, MBWLD is appropriate for face recognition, and the recognition performance is much well improved especially in illumination and age changes.

Introduction

Face recognition systems are frequently challenged by environmental variations, most notably changes in illumination and the natural aging process. Traditional descriptors like Local Binary Patterns (LBP) and the Weber Local Descriptor (WLD) primarily operate at the pixel level. While effective at capturing micro-patterns, these "too local" approaches are highly sensitive to noise and often fail to represent broader facial structures, known as macro-patterns.

To address these limitations, this research introduces the Multi-scale Block Weber Local Descriptor (MBWLD). By analyzing regional averages rather than individual pixels, MBWLD aims to provide a representation that is both descriptive enough for fine detail and robust enough to ignore environmental noise.

Methodology

The MBWLD generation process follows a three-stage pipeline designed to capture both local and global facial features:

  1. Multi-scale Block (MB) Filtering: The input image is processed using square sub-regions (blocks) of varying scales (e.g., $3 \times 3$ to $21 \times 21$). The average gray-value for each block is calculated, effectively smoothing out pixel-level noise and capturing larger structural information.
  2. Weber Local Descriptor (WLD) Extraction: The system applies Weber’s Law logic to these block averages. This consists of two main components:
  3. Differential Excitation ($\xi$): This calculates the ratio between the relative change of a block against its neighbors and the block's own intensity, mimicking human perception's sensitivity to relative changes.
  4. Orientation ($\theta$): Gradient directions are determined using Sobel-like filters to capture edge information.
  5. Histogram Construction: These components are combined into a final histogram that serves as a robust "fingerprint" for the face.
The MBWLD pipeline, illustrating the integration of multi-scale block filtering with Weber Local Descriptor components.

By applying WLD mathematics to multi-scale block averages, the descriptor successfully integrates the "macro" awareness of regional blocks with the "micro" robustness of Weber’s Law.

Experimental Results

The proposed method was evaluated using the FERET database, a standard benchmark containing over 14,000 images. Testing focused on four specific subsets: expression changes (fb), illumination changes (fc), and two sets representing age changes (dup I and the more challenging dup II).

The MBWLD was compared against standard LBP, Multi-scale Block LBP (MBLBP), and the original WLD. Key findings include:

  • Illumination Robustness: MBWLD significantly outperformed MBLBP, which typically struggles under lighting variations. The WLD component effectively "rescues" the block-based approach from its inherent sensitivity to light.
  • Aging Performance: On the difficult "dup II" dataset, MBWLD achieved the highest recognition rates. As shown in the performance curves, MBWLD consistently maintains a lead over other descriptors as the rank (the number of top candidates considered) increases.
  • Expression Invariance: While all tested descriptors performed well (exceeding 95% accuracy), MBWLD remained highly competitive with state-of-the-art methods.
Comparison of recognition rates across different ranks for LBP, MBLBP, WLD, and the proposed MBWLD on FERET subsets.

The researchers also noted that the differential excitation histogram for MBWLD is more evenly distributed than that of standard WLD, indicating a more descriptive feature set that is better at distinguishing between different individuals.

Conclusion

The Multi-scale Block Weber Local Descriptor (MBWLD) offers a significant advancement in robust face recognition. By bridging the gap between pixel-level detail and regional structure, it provides superior performance in the face of illumination and aging challenges. Future work will focus on improving computational efficiency and exploring entirely new local features to further enhance recognition accuracy.

Summary updated

Abstract

In this paper, we provide a novel local descriptor Multi-scale Block Weber Local Descriptor (MBWLD) for face recognition. It is a LBP-based descriptor which combines MBLBP with WLD for integrating both their advantages. Firstly, Multi-scale block filters is used on the face image to calculate the average gray-values of sub-regions. Secondly, differential excitation and orientation of WLD is extracted using the average neighbors and center. MBWLD extracts both the micro and macro patterns from face image, and becomes more descriptive and robust to face representation. The feasibility and availability are validated by comparing with the standard LBP, MBLBP and WLD on FERET database in expression, illumination and age changes. The experimental result shows that, MBWLD is appropriate for face recognition, and the recognition performance is much well improved especially in illumination and age changes.
BibTeX
@misc{pu2015multi,
  author = {Weilan Pu and Xiaohong Zhang},
  title = {{Multi-scale Block Weber Local Descriptor for Face Recognition}},
  year = {2015},
  doi = {10.2991/ameii-15.2015.292},
}
Figure