Abstract
Color constancy is the ability to measure colors of objects independent of the light source color. Various methods have been proposed to handle this problem. Most of them depend on the statistical distributions of the pixel values. Recent studies show that incorporation image derivatives are more effective than the direct use of pixel values. Based on this idea, a novel edge-based color constancy algorithm using support vector regression (SVR) is proposed. Contrary to existing SVR color constancy algorithm, which is computed from the zero-order structure of images, our method is based on the higher-order structure of images. The experimental results show that our algorithm is more effective than the zero-order SVR color constancy methods.
Introduction
Color constancy is a critical task in computer vision, enabling systems to perceive the "true" color of objects regardless of varying illumination conditions. Traditionally, algorithms have relied on "zero-order" statistics—the distribution of raw pixel values. However, these methods often struggle with diffuse lighting and color offsets. This paper introduces Edge-Based Color Constancy via Support Vector Regression (E-SVR), a method that estimates scene illumination by analyzing image derivatives (edges) rather than raw pixel values. By training a Support Vector Regressor on binarized chromaticity histograms of edges, the authors achieve significant improvements in accuracy over traditional pixel-based methods.
Methodology
The proposed E-SVR pipeline transforms an image into a feature vector suitable for regression through several key steps:
Edge Extraction and Smoothing
First, the algorithm calculates partial derivatives ($f_x, f_y$) to identify reflectance changes and computes the gradient magnitude to create an "edge image." A Gaussian filter is then applied to this edge image. The authors identified that a smoothing parameter ($\sigma = 7$) is critical for optimal performance, as it effectively balances noise reduction with feature preservation.
Chromaticity Transformation and Binarized Histograms
To remove intensity shading, which depends on geometry rather than light color, the edge data is converted into $(r, g)$ chromaticity space. Instead of using standard histograms, the authors employ binarized histograms ($n \times n$ bins). Each bin is marked '1' if a specific chromaticity is present in the edges and '0' otherwise. This binarization reduces input dimensionality and focuses the model on the presence of certain colors at edges, making it robust to object size.
Support Vector Regression (SVR)
An SVR with a Radial Basis Function (RBF) kernel is used to map these edge-based features to an accurate estimate of the light source. SVR was chosen for its ability to achieve a global optimum and its straightforward training process compared to other machine learning models of the time.
Experimental Results
The model was evaluated using Barnard’s dataset, a standard benchmark consisting of 321 real-world images across 30 different scenes and 11 light sources. A leave-one-out cross-validation procedure was employed to ensure generalizability.
The results, measured by median angular error (where lower is better), demonstrate the superiority of the edge-based approach:
- Standard SVR (Zero-order): 4.65°
- Proposed E-SVR (Higher-order): 3.77° (~18.9% improvement)
Compared to other statistical methods, E-SVR showed a consistent 16% to 19% improvement when moving from pixel-based to edge-based features. While specialized methods like Edge-based Gamut mapping performed slightly better in absolute terms (2.60°), E-SVR significantly outperformed the standard learning-based SVR approach.
Discussion and Conclusion
The core insight of this work is that the distribution of colors at the edges of objects provides a cleaner signal for illumination than surface colors themselves. Edges are naturally more invariant to certain lighting offsets. However, the model currently assumes a single, uniform light source and is sensitive to the Gaussian smoothing parameter ($\sigma$).
Future work could involve extending learning methods to even higher-order structures beyond first-order derivatives and adapting the model for complex scenes with multiple light sources.
Summary updated