Abstract
Face recognition performance is significantly degraded by complex illumination conditions, particularly sharp shadow edges that violate the assumption of slowly varying light. This paper proposes a dual-model illumination estimation approach that combines a local maximum model for smooth regions with an edge-aware model for shadow boundaries. By fusing these models and applying adaptive anisotropic Gaussian filtering, the method extracts illumination invariants that preserve critical facial features. Experimental results on the Yale B database show that the proposed method achieves up to 100% recognition accuracy in challenging subsets, outperforming traditional techniques like Multiscale Retinex and Gradientfaces.
Introduction
The primary challenge in face recognition under varying lighting is the separation of the object's intrinsic reflectance ($R$) from the incident illumination ($I$). According to the Lambertian reflectance model, an image $F$ is the product of these two components: $F = R \times I$. To achieve illumination-invariant recognition, systems must accurately estimate and remove $I$.
Traditional methods often assume that illumination changes slowly across the face. While this holds for soft lighting, it fails in the presence of "light-shielding edges"—sharp shadow boundaries caused by harsh side-lighting. Standard filters like Gaussian or Multiscale Retinex (MSR) tend to blur these edges, leading to artifacts in the extracted invariants and reduced recognition accuracy. This paper introduces a method specifically designed to handle these sharp transitions.
Proposed Method
The proposed approach consists of three main stages to estimate the illumination map and extract the invariant reflectance.
1. Dual-Model Estimation
The system employs two distinct models to capture different lighting characteristics:
- Model I ($I_m$): Estimates illumination using the local maximum value within a small window. This is effective for regions with smooth or slowly changing light.
- Model II ($I_s$): A specialized calculation designed to accurately track the rapid intensity changes at shadow edges.
2. Illumination Fusion
To combine these models, the system calculates a global threshold ($t$) based on image statistics. It performs pixel-wise segmentation to decide whether to apply the smooth estimate (Model I) or the edge-aware estimate (Model II), ensuring that sharp transitions are preserved where they occur.
3. Refinement via Adaptive Filtering
The fused illumination map is refined using an adaptive anisotropic Gaussian filter. This step ensures spatial consistency and removes noise while respecting the structural boundaries identified during fusion. The final illumination invariant ($R$) is then obtained by dividing the original image by this refined estimate: $R = F / I$.
Experimental Setup
The method was evaluated using the Combined Yale B database, which includes 38 subjects under 64 different lighting conditions. Images were resized to 100x100 pixels and categorized into five subsets (S1 to S5) based on the severity of the light's incident angle, with S5 representing the most extreme and difficult conditions.
Results and Discussion
The proposed method demonstrates significant improvements in both visual fidelity and quantitative recognition rates.
Recognition Accuracy
The method consistently outperformed established benchmarks such as MSR, Gradientfaces, and Tan & Triggs (TT). In the most difficult lighting category (Subset 5), the proposed method achieved a 100% recognition rate when using S5 as training data. More importantly, when training on the easiest set (S1) and testing on the hardest (S5), it maintained a high accuracy of 98.06%, compared to 94.74% for Gradientfaces and 94.04% for MSR.
Parameter Sensitivity
Analysis of the window size ($\Omega$) revealed that a small $3 \times 3$ window provides the optimal balance for local illumination estimation. Larger windows tend to over-smooth the estimate, losing the benefits of the dual-model approach at shadow edges.
Conclusion
By explicitly modeling the behavior of light at occlusion edges, this dual-model approach effectively addresses the limitations of traditional illumination-invariant methods. While the results on controlled datasets like Yale B are excellent, future work should focus on optimizing computational efficiency for real-time applications and testing the model's robustness in unconstrained "in-the-wild" environments.
Summary updated