Deep Single Image Defocus Deblurring via Gaussian Kernel Mixture Learning.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Defocus blur is a common photographic artifact that degrades image quality.
• Spatially-varying blur in defocus images presents a significant challenge for restoration.
• Existing deblurring methods struggle with efficiency and accuracy for large, complex blurs.

Purpose of the Study:

• To develop an end-to-end deep learning approach for single-image defocus deblurring.
• To accurately model and remove spatially-varying defocus blur.
• To achieve high-quality image restoration with improved computational efficiency.

Main Methods:

• Proposed a pixel-wise Gaussian kernel mixture (GKM) model to parameterize spatially-varying defocus point spread functions (PSFs).
• Introduced a grouped GKM (GGKM) model to enhance modeling accuracy efficiently by decoupling coefficients.
• Developed GGKMNet, a deep neural network unrolling a fixed-point iteration of GGKM-based deblurring, using a scale-recurrent architecture for coarse-to-fine coefficient estimation.

Main Results:

• GGKMNet successfully recovers all-in-focus images from single defocused images.
• The model demonstrates superior restoration quality compared to existing defocus deblurring methods across five benchmark datasets.
• GGKMNet exhibits reduced model complexity and enhanced computational efficiency.

Conclusions:

• The proposed GGKMNet effectively addresses the challenges of spatially-varying defocus blur using an innovative deep learning framework.
• The method offers a computationally efficient and accurate solution for single-image defocus deblurring.
• GGKMNet represents a significant advancement in image restoration for photographic applications.