Abstract
Spatial structure in a color image can be represented using correlation functions defined within and between sensor bands. Using a linear model for surface spectral reflectance with the same number of parameters as the number of classes of photoreceptors, we show that illumination changes correspond to linear transformations of a surface correlation matrix. From this relationship, we derive a distance function for comparing sets of spatial correlation functions that can be used for illumination-invariant recognition. We demonstrate using a large body of experiments that this distance function can be used for accurate texture classification in the presence of large changes in illumination spectral distribution.
Introduction
A fundamental challenge in computer vision is "color constancy"—the ability to recognize a surface's properties regardless of the light source's spectral composition. For example, a texture might appear drastically different when viewed under direct sunlight versus incandescent light or colored filters. Standard recognition algorithms often fail when pixel values shift due to these illumination changes. This paper proposes a mathematical framework for illumination-invariant recognition by modeling how changes in lighting affect the spatial correlation of color bands.
Methodology
The authors' approach is built on two pillars: spatial correlation functions and linear reflectance models.
Spatial Correlation Functions
Instead of analyzing isolated pixels, the model uses spatial correlation functions to capture the relationship between color values at different locations. By considering both intra-band (e.g., Red-Red) and inter-band (e.g., Red-Green) correlations, the system creates a robust representation of the texture's spatial and spectral structure.
Linear Reflectance Models and Subspace Invariance
The core of the method assumes that surface spectral reflectance can be approximated by a finite-dimensional linear model. In a typical three-sensor (RGB) system, the authors show that a change in illumination acts as a linear transformation on the spatial correlation matrix. Specifically, if $\tilde{C}$ is the correlation matrix under a reference light and $C$ is the matrix under a new light, then $C = \tilde{C}L$, where $L$ is a transformation matrix.
To classify a new texture, the system calculates a distance $D$ using Singular Value Decomposition (SVD). By projecting the test texture's correlation vectors onto an orthonormal basis derived from the database texture, the system identifies matches based on the projection error. This error remains small even if the lighting conditions are significantly different, as the underlying subspace spanned by the correlation functions remains invariant.
Experimental Results
The method was tested on a dataset of 20 diverse textures, including natural materials like sand and trees, as well as man-made items like carpets and wallpaper. Each texture was imaged under four different lighting conditions: white (for the database), yellow, red, and green.
Performance and Discriminative Power
The proposed distance function $D$ achieved a 100% success rate, correctly identifying all 60 test images across the yellow, red, and green lighting conditions. In comparison, a standard Euclidean distance metric ($D'$) that ignores illumination changes only achieved 55% accuracy (33 out of 60).
As shown in the results, the distance $D$ for the correct texture match is typically orders of magnitude smaller than for incorrect textures, demonstrating the high discriminative power and robustness of the subspace-based approach.
Conclusion
By combining linear models of reflectance with spatial correlation analysis, the authors developed a method for texture classification that is essentially "blind" to the color of the light source. While the current study used a relatively small dataset, the perfect accuracy across extreme lighting shifts suggests that subspace invariance is a powerful tool for robust computer vision. Future work may involve integrating this method with geometric invariance to handle changes in scale and orientation.
Summary updated