Blind image deblurring using spectral properties of convolution operators.

Area of Science:

• Image processing
• Computer vision
• Signal processing

Background:

• Blind deconvolution aims to restore sharp images from blurry ones when the blur kernel is unknown.
• Existing methods require effective regularization for both the sharp image and blur kernel due to the ill-conditioned nature of the problem.
• Regularizing the blur kernel effectively remains a challenge in blind deconvolution.

Purpose of the Study:

• To develop a new approach for blind deconvolution by effectively regularizing the unknown blur kernel.
• To demonstrate that blurry images contain sufficient information for blur kernel estimation.
• To improve the accuracy and performance of image deblurring techniques.

Main Methods:

• Analyzing the spectral changes of an image before and after blurring to retrieve the blur kernel.
• Establishing a novel convex kernel regularizer that depends solely on the blurry image.
• Integrating the proposed regularizer with existing non-blind deconvolution techniques.

Main Results:

• The proposed method successfully retrieves blur kernel information from the blurry image spectrum.
• A new convex kernel regularizer is introduced, dependent only on the blurry image.
• Combining the regularizer with non-blind deconvolution significantly enhances deblurring performance on synthetic and real images.

Conclusions:

• The blurry image itself contains valuable information for estimating the blur kernel.
• The developed convex kernel regularizer provides a robust and effective way to constrain the blur kernel.
• The proposed blind deconvolution approach offers significant improvements over existing methods, particularly when the original image possesses sufficient sharpness.