•  
  •  
 

Abstract

Image denoising plays a vital role in enhancing visual quality by effectively suppressing noise while retaining critical image structures and textures. Traditional Block-Matching and 3D (BM3D) Filtering techniques, although widely adopted, often encounter challenges in achieving an optimal trade-off between noise reduction and feature preservation due to limitations in fixed-thresholding strategies and suboptimal block matching. To address these shortcomings, this study introduces a novel Cosine Adaptive BM3D (CA-BM3D) approach, which integrates cosine similarity for more accurate block matching and incorporates adaptive thresholding to enhance denoising efficiency. The proposed method was evaluated on six standard 8-bit grayscale images such as Leena (512*512), Barbara (512*512), Zelda (512*512), peppers (512*512), Cameraman (512*512), House (512*512) contaminated with Additive White Gaussian Noise (AWGN) Results demonstrate that the Cosine Adaptive BM3D algorithm consistently outperforms conventional BM3D across all test cases, achieving an average PSNR of 31.28 dB and SSIM of 0.678. Notably, for the ``Leena'' image, the cosine-based approach attained a PSNR of 31.42 dB and SSIM of 0.8921, surpassing alternative distance metrics such as Euclidean (30.55 dB, 0.8782), Manhattan (29.96 dB, 0.8657), Jaccard (28.78 dB, 0.8425), and Minkowski (30.02 dB, 0.8691). Similar performance gains were observed for the other images: ``Barbara'' (30.33 dB, 0.8856), ``Zelda'' (32.15 dB, 0.9012), ``Peppers'' (31.78 dB, 0.8967), ``Cameraman'' (30.02 dB, 0.8823), and ``House'' (32.89 dB, 0.9105). These findings substantiate the effectiveness of the cosine similarity measure in enhancing both noise suppression and structural fidelity in the denoising process.

Share

COinS