SSIM Compliant Modeling Framework With Denoising and Deblurring Applications.

Area of Science:

• Image Processing and Computer Vision
• Signal Processing
• Machine Learning

Background:

• Mean square error is perceptually inadequate for image quality assessment (IQA).
• Structural Similarity Index (SSIM) is a popular perceptual measure but its non-convexity poses challenges for model-based applications.
• Existing IQA methods struggle with the non-convex nature of SSIM in applications like denoising and restoration.

Purpose of the Study:

• To address the non-convexity issues of SSIM in image processing.
• To develop a generalized convex framework for SSIM-based applications.
• To enhance image denoising and deblurring using a novel learning scheme.

Main Methods:

• Developed a generalized convex framework for SSIM.
• Integrated the framework into an alternative learning scheme for regularized linear models.
• Incorporated a dictionary learning module and sparsity prior.
• Introduced an iterative scheme based on noise statistics.

Main Results:

• The proposed framework effectively handles the non-convexity of SSIM.
• The alternative learning scheme with sparsity prior significantly improved denoising and deblurring.
• The iterative scheme further boosted performance by considering noise characteristics.
• Achieved competitive and superior performance on natural images compared to existing methods.

Conclusions:

• The generalized convex framework provides a robust solution for SSIM-based image processing tasks.
• The proposed learning scheme offers a powerful alternative for image restoration applications.
• The method demonstrates state-of-the-art performance, particularly in visual quality and SSIM metrics for natural images.