Unsharp Mask Guided Filtering.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Guided image filtering is crucial for structure transfer using a guidance image.
• Deep networks have advanced guided filters by learning coefficients, but simultaneous estimation causes artifacts.
• Classical unsharp masking uses a single coefficient for edge enhancement.

Purpose of the Study:

• To propose a simplified guided filter formulation that estimates a single coefficient.
• To introduce a successive guided filtering network for efficient and accurate image processing.
• To address halo artifacts and structure inconsistencies in current guided filters.

Main Methods:

• A novel guided filter formulation inspired by unsharp masking, estimating a single coefficient.
• Integration of a low-pass filter prior for explicit structure transfer.
• Development of a successive guided filtering network for multi-resolution outputs.

Main Results:

• The proposed single-coefficient formulation effectively reduces halo artifacts and improves structure consistency.
• The successive guided filtering network achieves state-of-the-art performance in upsampling, denoising, and cross-modality filtering.
• Demonstrated trade-off between accuracy and efficiency with multiple filtering results from one network.

Conclusions:

• The simplified guided filter formulation is more effective and efficient than previous methods.
• The proposed network advances image filtering tasks by enabling explicit structure transfer.
• This work offers a robust and versatile approach to guided image filtering.