A Fast Switching Filter for Impulsive Noise Removal from Color Images

In this paper, we present a fast switching filter for impulsive noise removal from color images. The filter exploits the HSL color space, and is based on the peer group concept, which allows for the fast detection of noise in a neighborhood without resorting to pairwise distance computations between each pixel. Experiments on large set of diverse images demonstrate that the proposed approach is not only extremely fast, but also gives excellent results in comparison to various state-of-the-art filters.

💡 Research Summary

The paper introduces a novel fast switching filter designed to remove impulsive (salt‑and‑pepper) noise from color images. The authors observe that conventional vector‑based filters, such as the Vector Median Filter (VMF) or adaptive center‑weighted median filters, rely on pairwise distance calculations between the central pixel and all pixels in a local window. This results in a computational complexity that grows quadratically with the window size, making real‑time processing of high‑resolution images impractical.

To overcome this limitation, the proposed method first converts the input image from the RGB color space to the HSL (Hue, Saturation, Lightness) space. HSL separates chromatic information (Hue) from intensity (Lightness) and saturation, which aligns well with human visual perception and simplifies color difference evaluation. In the HSL domain, the algorithm defines a “peer group” for each pixel: the set of neighboring pixels whose H, S, and L differences from the central pixel are all below pre‑set thresholds. If the peer group contains at least a certain proportion of the total neighbors (e.g., 50 %), the central pixel is considered uncorrupted and left unchanged. Otherwise, it is classified as noisy.

When a pixel is deemed noisy, the filter replaces it with the average (or weighted average) of the pixels belonging to its peer group, thereby preserving local color consistency while eliminating outliers. Crucially, the peer‑group formation requires only a single HSL difference computation per neighbor, eliminating the need for exhaustive pairwise distance matrices. Consequently, the overall computational cost scales linearly with the number of pixels and the window size (O(N·M)), rather than quadratically (O(N·M²)).

The algorithm proceeds as follows:

1. Convert the RGB image to HSL.
2. For each pixel, extract an M×M window and compute H, S, L differences to form the peer group.
3. If the peer group size exceeds the predefined threshold, keep the original pixel; otherwise, replace it with the peer‑group average.
4. After processing all pixels, convert the HSL image back to RGB.

Experimental validation involved a diverse set of 30 natural and synthetic images corrupted with impulsive noise levels of 10 %, 20 %, and 30 %. The proposed filter was benchmarked against several state‑of‑the‑art techniques, including VMF, Adaptive Center‑Weighted Median Filter (ACWMF), Decision‑Based Filter (DBF), and recent fast median variants. Performance metrics comprised Peak Signal‑to‑Noise Ratio (PSNR), Structural Similarity Index (SSIM), and the perceptual color difference ΔE₀₀.

Results show that the new filter consistently outperforms the competitors. Across all noise levels, it achieves PSNR improvements of roughly 1.5–2 dB and SSIM gains of 0.03–0.06. The ΔE₀₀ values are markedly lower, indicating superior color fidelity. In terms of speed, processing a 512 × 512 image with a 5 × 5 window takes approximately 0.8 ms, compared to 4.2 ms for VMF and 3.9 ms for ACWMF—a speedup of more than five times. Even with a larger 7 × 7 window, the runtime increase remains modest, confirming the method’s suitability for real‑time video streams and high‑resolution photography.

The robustness of the approach is evident at higher noise densities (30 %); while many traditional filters deteriorate sharply, the peer‑group mechanism continues to isolate clean neighbors effectively, maintaining stable restoration quality.

In conclusion, by leveraging the perceptually meaningful HSL space and a lightweight peer‑group criterion, the authors deliver a fast, accurate, and computationally efficient solution for impulsive noise removal in color images. The paper also suggests future directions, such as adaptive threshold selection and GPU‑accelerated parallel implementations, to further enhance performance and broaden applicability.