Blind Noisy Image Quality Assessment Using Sub-Band Kurtosis

Area of Science:

• Digital Image Processing
• Computer Vision
• Signal Processing

Background:

• Image noise significantly degrades perceptual quality and masks content.
• Statistical differences exist between clean and noisy images in wavelet coefficient distributions.
• Kurtosis measures the peakedness and tail weight of these distributions.

Purpose of the Study:

• To evaluate the efficacy of sample kurtosis of wavelet coefficients for image quality assessment.
• To develop an extreme learning machine model that maps kurtosis to perceptual quality scores.
• To assess the model's performance on various noise types and compare it with existing methods.

Main Methods:

• Utilized discrete wavelet transform (DWT) to analyze image coefficients.
• Calculated sample kurtosis of DWT coefficients as a key feature.
• Trained an extreme learning machine (ELM) model using kurtosis values.
• Tested the model on diverse noisy image databases (LIVE, CSIQ, TID2008, TID2013).

Main Results:

• The ELM model effectively learned the mapping from kurtosis to perceptual quality.
• The model demonstrated superior performance in evaluating noisy images compared to blind noise assessment models.
• Outperformed general-purpose blind and full-reference image quality assessment methods.

Conclusions:

• Sample kurtosis of wavelet coefficients is a potent feature for image quality assessment.
• The developed ELM-based model offers a robust and accurate solution for evaluating perceptual quality in noisy images.
• This approach provides a significant advancement over existing blind and full-reference quality assessment techniques.