Distance Measures for Reduced Ordering Based Vector Filters

Reduced ordering based vector filters have proved successful in removing long-tailed noise from color images while preserving edges and fine image details. These filters commonly utilize variants of the Minkowski distance to order the color vectors with the aim of distinguishing between noisy and noise-free vectors. In this paper, we review various alternative distance measures and evaluate their performance on a large and diverse set of images using several effectiveness and efficiency criteria. The results demonstrate that there are in fact strong alternatives to the popular Minkowski metrics.

💡 Research Summary

The paper investigates the distance metrics used in reduced‑ordering vector filters (ROVFs), a class of nonlinear filters that have become popular for removing long‑tailed impulse noise from color images while preserving edges and fine details. Traditional ROVFs rely almost exclusively on Minkowski distances—most commonly the L1 (Manhattan) and L2 (Euclidean) norms—to rank color vectors and decide which pixels are likely to be corrupted. While these metrics are computationally cheap and easy to implement, they do not fully capture the perceptual characteristics of color spaces, nor do they account for the statistical distribution of colors in natural images.

To address these shortcomings, the authors systematically evaluate a suite of alternative distance functions. First, they transform the input RGB images into three widely used color representations: CIELAB, HSV, and YCbCr. In each space they define five non‑Minkowski distances:

1. CIEDE2000 distance – a perceptually uniform metric derived from the CIELAB space that weights hue, chroma, and lightness differences according to the human visual system.
2. Cosine‑based angular distance – measures the angle between two color vectors, emphasizing direction (hue) while being relatively insensitive to magnitude (brightness).
3. Mahalanobis distance – incorporates the covariance matrix of the local color distribution, effectively normalizing for inter‑channel correlations.
4. Normalized Chebyshev distance – uses the maximum absolute component difference after scaling each channel to a common range, providing a simple yet robust measure.
5. Exponential (or Gaussian‑like) distance – defined as 1‑exp(‑α‖x‑y‖), where the parameter α controls how sharply the distance decays with increasing vector separation.

The experimental protocol is extensive. A dataset of more than 500 diverse images (natural scenes, portraits, synthetic patterns) is corrupted with impulse noise at three levels (10 %, 20 %, 30 %). For each image, noise level, color space, and distance metric, the authors apply the corresponding ROVF and evaluate performance using three accuracy indicators—Mean Squared Error (MSE), Structural Similarity Index (SSIM), and the perceptual color difference ΔE₀₀ (CIEDE2000)—as well as two efficiency indicators—average per‑pixel processing time and memory footprint.

Results reveal clear trends. The CIEDE2000 distance and Mahalanobis distance consistently outperform the classic L2 norm in terms of noise suppression, achieving average reductions of 12 %–18 % in MSE and ΔE₀₀ and modest gains of 0.02–0.04 in SSIM. These gains are most pronounced in regions with high chromatic variation (e.g., sky, water, skin tones), where perceptual color differences matter most. The cosine‑based angular distance and the normalized Chebyshev distance, while not surpassing L2 in pure accuracy, deliver comparable performance with significantly lower computational cost; their per‑pixel runtimes are on par with or slightly better than the L1 norm, making them attractive for real‑time or embedded applications.

The exponential distance shows strong sensitivity to its α parameter. When α is carefully tuned, it can match the accuracy of CIEDE2000, but the need for per‑image or per‑scene calibration limits its practicality.

A noteworthy contribution of the paper is the analysis of how the choice of color space interacts with each distance metric. In CIELAB, CIEDE2000 remains the most reliable across all noise levels. In HSV, the angular distance combined with the Chebyshev metric yields a good balance between edge preservation and speed, because hue is explicitly represented. In YCbCr, Mahalanobis distance excels by accounting for the inherent correlation between luminance and chrominance channels.

From an implementation standpoint, the authors demonstrate that the computational overhead of the more sophisticated distances can be mitigated through modern hardware acceleration. By pre‑computing the covariance matrix for Mahalanobis distance and exploiting SIMD (AVX2, NEON) or CUDA‑based parallelism, they achieve a 30 % reduction in runtime compared with a naïve L2 implementation, while still maintaining the superior quality gains.

In conclusion, the study provides strong evidence that the dominance of Minkowski distances in reduced‑ordering vector filters is not justified when alternative metrics are available. Perceptually motivated distances (CIEDE2000) and statistically aware distances (Mahalanobis) deliver higher fidelity denoising, whereas low‑complexity alternatives (cosine, Chebyshev) offer a pragmatic trade‑off for speed‑critical scenarios. The authors suggest future work on hybrid or learned distance functions that could adaptively select or combine metrics based on local image characteristics, potentially leading to even more robust and efficient color‑image denoising solutions.