$L_0$ -Regularized Intensity and Gradient Prior for Deblurring Text Images and Beyond

Area of Science:

• Computer Vision
• Image Processing
• Computational Imaging

Background:

• Text image deblurring is crucial for document analysis and recognition.
• Existing edge-based methods often rely on heuristic edge selection, limiting their robustness.
• Artifacts in deblurred images can degrade the quality of restored text.

Purpose of the Study:

• To develop a robust and efficient L0-regularized prior for text image deblurring.
• To improve kernel estimation accuracy by leveraging image intensity and gradient properties.
• To enhance the final latent image restoration by effectively removing artifacts.

Main Methods:

• A novel L0-regularized prior incorporating intensity and gradient information specific to text images.
• An efficient optimization algorithm for reliable intermediate results in kernel estimation.
• An artifact removal technique for the final latent image restoration step.

Main Results:

• The proposed algorithm achieves superior performance in text image deblurring compared to state-of-the-art methods.
• The method demonstrates robustness and effectiveness in handling complex scenes and low illumination.
• The algorithm successfully extends to non-uniform deblurring tasks.

Conclusions:

• The proposed L0-regularized prior offers a principled and effective approach to text image deblurring.
• The developed optimization and restoration techniques significantly improve deblurring quality.
• The algorithm's versatility makes it applicable to a broader range of image deblurring challenges.