Document Type : Original Article
Author
Department of Statistics, Payame Noor University, Tehran, Iran
Abstract
Noise removal in color images is challenging because effective denoising must suppress noise while preserving edges, textures, and inter-channel relationships. This paper develops a Weighted Bootstrap Mean (WBM) method based on eigenvector analysis of a local data matrix constructed from color values within each pixel neighborhood. The eigenvector-based analysis, referred to as EIG, serves as the baseline and characterizes local inter-channel correlation, distinguishes smooth regions from edges and textured areas, and adapts smoothing to local image structure. EIG is not a contribution of this work; it is reformulated to provide a self-contained description of the baseline and a direct basis for comparison. The main contribution is WBM, which extends EIG by replacing its single plug-in weighted mean with an ensemble of exponentially weighted bootstrap resamples of the local neighborhood. Samples closer to the local mean receive higher selection probabilities, reducing the influence of outliers and high-amplitude noise while retaining local information. Experiments on the Lenna and Flower images at σ = 10, 20, and 30 show that WBM consistently achieves the lowest MAE among CRF, FWD, and EIG. WBM also provides the highest PSNR in most cases. Compared with EIG, WBM consistently improves MAE and achieves higher PSNR on Flower. On Lenna at σ = 30, however, EIG yields a marginally higher PSNR, indicating that WBM’s principal advantage lies in robustness and detail preservation rather than uniform improvement across metrics. These findings demonstrate the potential of weighted bootstrap aggregation as a robust extension of eigenvector-based local denoising.
Highlights
-
Develops a Weighted Bootstrap Mean (WBM) method for color image denoising.
-
Integrates eigenvector-based local structure analysis with weighted bootstrap aggregation.
-
Evaluates denoising performance on Lenna and Flower images at three noise levels.
-
Achieves the lowest mean absolute error (MAE) among all compared methods.
-
Improves PSNR in most cases while enhancing robustness to noise and preserving image details.
Keywords