Scholarlink

<p>Impulse noise removal in color images remains challenging, particularly in regions with strong pixel correlations where conventional detection strategies fail. This work presents a statistically driven denoising framework that employs prediction-error analysis for robust noise localization and adaptive restoration. Unlike existing methods relying on direct intensity comparisons, the proposed approach models directional prediction-error statistics and incorporates noise density awareness through a lightweight CNN classifier. This enables selective restoration that preserves structural and chromatic consistency under varying noise conditions. Extensive experiments on natural and medical image datasets demonstrate superior quantitative and perceptual performance, achieving PSNR gains of up to 1.8 dB and SSIM improvements of 0.02–0.05 over state-of-the-art filters, underscoring both the methodological novelty and robustness of the proposed framework.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CNN-assisted statistical prediction-error analysis for impulse noise removal in color images

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Keisham Pritamdas

摘要

Impulse noise removal in color images remains challenging, particularly in regions with strong pixel correlations where conventional detection strategies fail. This work presents a statistically driven denoising framework that employs prediction-error analysis for robust noise localization and adaptive restoration. Unlike existing methods relying on direct intensity comparisons, the proposed approach models directional prediction-error statistics and incorporates noise density awareness through a lightweight CNN classifier. This enables selective restoration that preserves structural and chromatic consistency under varying noise conditions. Extensive experiments on natural and medical image datasets demonstrate superior quantitative and perceptual performance, achieving PSNR gains of up to 1.8 dB and SSIM improvements of 0.02–0.05 over state-of-the-art filters, underscoring both the methodological novelty and robustness of the proposed framework.