Abstract
Computational color constancy is a classical problem in computer vision, aiming to estimate the true colors of a scene by discounting the effect of the illuminant. This is an under-constrained problem typically solved through various physical or statistical constraints. In this paper, we propose a novel hypothesis: images captured under the same illumination share similar statistical features. Based on this, we present the Color Constancy Based on Image Similarity (CCBIS) algorithm. Our method utilizes low-dimensional Minkowski-norm features and a similarity-based search against a database of known illuminants. Experimental results demonstrate that CCBIS is highly effective, achieving lower error rates than traditional methods while using significantly fewer feature dimensions.
1. Introduction
The appearance of an image is a product of object reflectance, sensor characteristics, and the spectral power distribution of the light source. Color constancy is the ability to recover the true reflectance of objects regardless of the illumination. For computers, this is a challenging "under-constrained" problem because multiple combinations of light and reflectance can produce the same pixel values.
Existing statistical methods often address this by analyzing the distribution of chromaticities in an image. These methods frequently represent images using high-dimensional feature matrices—for instance, a $25 \times 25$ grid in $(r, g)$ chromaticity space—to track which colors are present in a scene.
While these high-dimensional approaches can be accurate, they are computationally intensive. This paper introduces a more efficient alternative based on the hypothesis that images with similar structural and color statistics are likely captured under similar lighting conditions.
2. The CCBIS Algorithm
The proposed CCBIS algorithm estimates illumination by comparing a query image's features to a database of images with known light sources.
2.1 Preprocessing and Segmentation
The pipeline begins with preprocessing to remove dark pixels, which are prone to noise and provide unreliable data about the illuminant. The image is then segmented into regions using a region-growing method. This step groups pixels with similar chromaticity, simplifying the scene into a set of representative color patches.
2.2 Feature Extraction
A key contribution of this work is the use of a compact $3 \times 3$ feature matrix. Instead of a massive color histogram, the algorithm calculates the Minkowski norm of the average colors of the segmented regions. By employing different moments ($p=1, 6, \text{ and } \infty$), the algorithm captures a range of statistical information—from the global mean to the brightest highlights—in just 9 dimensions.
2.3 Similarity Search and Estimation
Similarity is defined by the angular distance between feature vectors. The algorithm identifies the $n$ most similar images in the database and estimates the query image's illumination as a weighted sum of their known light sources. This approach prioritizes the direction of color vectors in RGB space, which is more indicative of the illuminant's color than its intensity.
3. Experimental Results
The CCBIS algorithm was tested on a benchmark dataset of 429 images. A subset of 108 images formed the reference database, while 321 images across 30 scenes were used for evaluation.
Compared to the classic Color-by-Correlation method, CCBIS demonstrated significant improvements:
- Efficiency: CCBIS uses a 9-dimensional feature space, a massive reduction from the 625 dimensions used by Color-by-Correlation.
- Accuracy: CCBIS achieved a Mean Angular Error of 7.4, outperforming Color-by-Correlation's error of 9.0.
- Qualitative Quality: Visual comparisons showed that images corrected by CCBIS were nearly indistinguishable from those corrected using ground-truth illumination data.
4. Conclusion
This paper presents a novel, similarity-based approach to color constancy. By leveraging low-dimensional Minkowski-norm features, the CCBIS algorithm provides a computationally efficient and accurate solution. The success of this method suggests that high-dimensional color distributions are not always necessary for effective illumination estimation. Future work will explore the use of relevance feedback to further enhance the algorithm's performance.
Reference: Li, B., Xu, D., Wang, J. H., & Lu, R. (2008). Color Constancy Based on Image Similarity. IEICE Transactions on Information and Systems, E91-D(2), 375-378.
Summary updated