Performance Comparison of Five Spatial Filtering Algorithms for Salt-and-Pepper Noise Removal

Abstract

This work presents a rigorous quantitative evaluation of five core spatial filters for the mitigation of salt-and-pepper impulse noise, namely mean, median, maximum, minimum and order-statistic filters (with k=7), using MATLAB simulation. Three well-established standard greyscale benchmark images form the test dataset, with Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) and Signal-to-Noise Ratio (SNR) adopted as quantitative performance metrics. The bulk of published research focuses on refining single filter architectures, yet few studies offer holistic cross-algorithm testing within identical experimental constraints. To resolve this identified research gap, this paper quantifies operational use-cases for each classic spatial filter and develops measurable selection benchmarks for industrial image processing workflows, with direct relevance to clinical medical imaging, continuous surveillance capture and embedded vision hardware. Empirical testing confirms that the median filter achieves optimal noise reduction at a noise density of 0.1, recording average scores of 30.30 dB PSNR, 0.989 SSIM and 0.97 SNR. The denoising performance of the k=7 order-statistic filter falls midway between the median filter and the two extremum variants. Maximum and minimum filters produce negative SNR outputs, confirming their unsuitability for mixed dual-polarity salt-and-pepper noise. Adjusting the integer parameter k allows the order-statistic filter to adapt to differing noise polarities; peak performance of 26.58 dB PSNR occurs at k=5, with rapid degradation observed as k diverges from the median index.