Noise-Adaptive Non-Blind Image Deblurring.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image deblurring is crucial for applications facing noise and blur, common in low-light or high-speed imaging.
• Classical deconvolution methods struggle with arbitrary noise and complex blur kernels.
• Deep neural networks (DNNs) offer potential but require careful parameter optimization.

Purpose of the Study:

• To develop an advanced non-blind image deblurring method robust to arbitrary noise and blur.
• To improve DNN-based image enhancement through joint optimization of parameters and weights.
• To introduce a novel two-step DNN approach for enhanced deblurring and artifact removal.

Main Methods:

• A two-step deblurring system utilizing two DNNs.
• A novel RegParamNet architecture for estimating deconvolution regularization parameters.
• Joint optimization of regularization parameters and network weights for image enhancement.

Main Results:

• The proposed system effectively handles noise across a three orders of magnitude range (0.01-10.0).
• It performs well with a wide spectrum of 1D and 2D Gaussian blur kernels.
• Significantly outperforms several leading state-of-the-art deblurring approaches.

Conclusions:

• The novel DNN-based method provides superior image deblurring performance.
• The RegParamNet architecture enables robust regularization parameter estimation.
• The two-step approach effectively enhances image quality and removes residual artifacts.