The SURE-LET approach to image denoising.

Area of Science:

• Signal Processing
• Image Processing
• Computer Vision

Background:

• Traditional image denoising algorithms often rely on statistical models of the original image, which can be limiting.
• Minimizing noise in transformed domains can lead to suboptimal results if not carefully managed.

Purpose of the Study:

• To develop a new image denoising approach that bypasses the need for statistical image modeling.
• To introduce a method for performing denoising optimization in the image domain, even when processing occurs in a transformed domain.

Main Methods:

• Utilizing Stein's unbiased risk estimate (SURE) for mean squared error minimization directly in the image domain.
• Employing the linear expansion of thresholds (LET) principle to express the denoising process as a linear combination of elementary processes.
• Solving a linear system of equations derived from the SURE and LET principles for efficient denoising.

Main Results:

• Demonstrated that image-domain minimization is superior to subband minimization for tight frame transforms.
• Achieved competitive denoising results using a simple threshold on undecimated Haar wavelet coefficients, optimized via image-domain SURE.
• Showcased the efficiency and speed of the proposed SURE-LET based denoising algorithm.

Conclusions:

• The proposed SURE-LET framework offers a powerful and flexible approach to image denoising.
• The method's ability to perform optimization in the image domain, independent of statistical priors, is a significant advantage.
• The SURE-LET principle shows substantial potential for developing highly effective and computationally efficient image denoising algorithms.