Understanding Symmetric Smoothing Filters: A Gaussian Mixture Model Perspective.

Area of Science:

• Computer Vision
• Image Processing
• Statistical Learning

Background:

• Patch-based image denoising often uses smoothing filters.
• These filters require row normalization for consistent application.
• Performance gains are observed when column normalization precedes row normalization.

Purpose of the Study:

• To understand the performance gain from pre-column normalization in image denoising.
• To establish a statistical learning interpretation of the Sinkhorn-Knopp balancing algorithm.
• To develop a novel, improved image denoising algorithm.

Main Methods:

• Analyzing the Sinkhorn-Knopp algorithm from a statistical learning perspective.
• Demonstrating the equivalence between Sinkhorn-Knopp and an expectation-maximization (EM) algorithm for Gaussian mixture models.
• Developing the Gaussian mixture model symmetric smoothing filter (GSF) based on this correspondence.

Main Results:

• Sinkhorn-Knopp is shown to be equivalent to an EM algorithm for learning Gaussian mixture models.
• A geometrical interpretation of the symmetrization process is provided.
• The proposed GSF algorithm demonstrates superior performance over existing smoothing filters.

Conclusions:

• The performance gain in denoising is explained through the lens of statistical learning and Gaussian mixture models.
• GSF offers a generalized and effective approach to image denoising.
• GSF achieves competitive performance compared to state-of-the-art denoising methods.