Exploiting Non-Local Priors via Self-Convolution for Highly-Efficient Image Restoration.

Area of Science:

• Computational imaging
• Image processing
• Computer vision

Background:

• Effective priors are crucial for solving ill-posed inverse problems in image processing.
• Non-local similarity methods group similar patches for image modeling, achieving state-of-the-art results but are computationally expensive due to inefficient block matching.
• Existing non-local algorithms are time-consuming, primarily because of the extensive computation required for block matching.

Purpose of the Study:

• To propose a novel Self-Convolution operator for exploiting image non-local properties in a unified framework.
• To demonstrate that Self-Convolution can generalize existing non-local modeling methods and achieve comparable results with reduced computational cost.
• To introduce an efficient multi-modality image restoration scheme using Self-Convolution.

Main Methods:

• Developed a novel Self-Convolution operator to unify and enhance non-local image modeling.
• Utilized Fast Fourier Transform (FFT) for efficient implementation of the Self-Convolution operator.
• Proposed an online multi-modality image restoration scheme leveraging the Self-Convolution operator.

Main Results:

• Self-Convolution with FFT implementation accelerates popular non-local image restoration algorithms by two to nine times.
• The proposed online multi-modality image restoration scheme demonstrates superior denoising performance compared to existing methods.
• Achieved significant improvements in both efficiency and effectiveness for RGB-NIR image restoration.

Conclusions:

• Self-Convolution offers a computationally efficient alternative for non-local image modeling in image restoration.
• The proposed method significantly speeds up block matching, a bottleneck in traditional non-local algorithms.
• The Self-Convolution based multi-modality image restoration scheme provides state-of-the-art results for RGB-NIR images.