Edge-Aware BMA Filters.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Edge-aware filters are crucial for various computer vision and image processing tasks.
• Existing methods often lack a unified theoretical framework for filter development.

Purpose of the Study:

• To propose a principled approach for developing novel edge-aware filters.
• To introduce Bayesian Model Averaging (BMA) as a core principle for filter design.
• To enhance image processing by improving edge preservation and computational efficiency.

Main Methods:

• Formulating pixel filtering as an optimal estimation problem within local patches.
• Employing Bayesian Model Averaging (BMA) to combine multiple pixel estimates.
• Developing a family of BMA filters by varying cost functions and log-likelihood/log-prior functions.
• Introducing a BMA guided filter as an extension of the guided filter.

Main Results:

• BMA filters achieve smoothing comparable to state-of-the-art edge-aware filters.
• Two BMA filters demonstrate computational efficiency on par with the guided filter.
• The BMA guided filter exhibits superior sharp edge preservation compared to the standard guided filter.
• Filtered images from the BMA guided filter show similarity to results from clustering processes.

Conclusions:

• The proposed principled approach effectively generates versatile and efficient edge-aware filters.
• BMA filters represent a significant advancement in image processing, balancing smoothing and edge preservation.
• The BMA guided filter offers a novel and improved alternative to existing guided filters, with unique clustering-like properties.