Quantitative and Qualitative Evaluation of NLM and Wavelet Methods in Image Enhancement

TL;DR

This paper compares Non-local Means and Wavelet Thresholding for image denoising, analyzing their effectiveness across multiple datasets and noise types using various quality metrics to determine their strengths and limitations.

Contribution

It provides a comprehensive evaluation of NLM and Wavelet Thresholding methods across diverse datasets and noise conditions, highlighting their respective advantages and shortcomings.

Findings

NLM performs better in visual quality.

Wavelet Thresholding excels in frequency domain metrics.

Both methods are less effective on complex distortions like Dirty Lens.

Abstract

This paper presents a comprehensive analysis of image denoising techniques, primarily focusing on Non-local Means (NLM) and Daubechies Soft Wavelet Thresholding, and their efficacy across various datasets. These methods are applied to the CURE-OR, CURE-TSD, CURE-TSR, SSID, and Set-12 datasets, followed by an evaluation using Image Quality Assessment (IQA) metrics PSNR, SSIM, CW-SSIM, UNIQUE, MS-UNIQUE, CSV, and SUMMER. The results indicate that NLM and Wavelet Thresholding perform optimally on Set12 and SIDD datasets, attributed to their ability to effectively handle general additive and multiplicative noise masks. However, their performance on CURE datasets is limited due to the presence of complex distortions like Dirty Lens and Codec Error, which these methods are not well-suited to address. Analysis between NLM and Wavelet Thresholding shows that while NLM generally offers superior visual quality, Wavelet Thresholding excels in specific IQA metrics, particularly SUMMER, due to its enhancement in the frequency domain as opposed to NLM's spatial domain approach.