Learning Iteration-wise Generalized Shrinkage-Thresholding Operators for Blind Deconvolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Maximum a posteriori (MAP)-based blind deconvolution relies on salient edge selection and time-varying regularization.
• Existing methods often require handcrafted regularizers and parameter tuning, limiting edge enhancement.
• Current regularizers smooth images but do not effectively sharpen salient edges.

Purpose of the Study:

• To propose iteration-wise ℓp-norm regularizers within the MAP framework for improved blind deconvolution.
• To develop a data-driven strategy for blur kernel estimation, addressing limitations of handcrafted approaches.
• To enhance salient edges while suppressing trivial details and enabling time-varying regularization.

Main Methods:

• Extension of the generalized shrinkage-thresholding (GST) operator for ℓp-norm minimization with negative p values.
• Dynamical salient edge selection and time-varying regularization through iteration-wise GST parameter specification.
• A discriminative learning approach to learn GST operators from training data, coupled with a multi-scale scheme for efficiency.

Main Results:

• Negative p values effectively estimate the coarse blur kernel shape in early stages.
• Learned GST operators generalize well to different datasets and real-world blurry images.
• The proposed method outperforms state-of-the-art techniques in quantitative metrics and visual quality, with significantly improved speed.

Conclusions:

• The proposed iteration-wise ℓp-norm regularizers with data-driven learning offer a superior approach to blind deconvolution.
• This method effectively sharpens salient edges and provides accurate blur kernel estimation.
• The approach achieves state-of-the-art deblurring performance efficiently and generalizes well.