Laplacian Gradient Consistency Prior for Flash Guided Non-Flash Image Denoising.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Flash guided non-flash image denoising faces challenges in modeling cross-modality consistency.
• Existing pixel-level consistency models can cause blurred edges.

Purpose of the Study:

• To propose a novel Laplacian gradient consistency (LGC) model for flash guided non-flash image denoising.
• To develop an interpretable deep network (LGCNet) based on the LGC model.

Main Methods:

• Investigated the modality gap between flash and non-flash images in the gradient domain.
• Established the LGC model based on the finding that the modality gap follows a Laplacian distribution.
• Designed LGCNet, a deep network whose components align with the LGC model's solution.

Main Results:

• The LGC model exhibits faster convergence and higher denoising accuracy than pixel consistency models.
• LGCNet demonstrates superior quantitative and qualitative denoising performance compared to state-of-the-art methods on multiple datasets.
• Visualization of intermediate features confirms the effectiveness of the Laplacian gradient consistency prior.

Conclusions:

• The proposed LGC model and LGCNet offer an effective and interpretable solution for flash guided non-flash image denoising.
• Laplacian distribution in the gradient domain is a powerful prior for cross-modality image consistency.