Partial Deconvolution With Inaccurate Blur Kernel.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Non-blind deconvolution methods often assume perfect blur kernel estimation, leading to artifacts when errors occur.
• Blind deconvolution, while advanced, still results in inevitable blur kernel estimation errors.
• These errors introduce severe artifacts like ringing and distortions in deblurred images.

Purpose of the Study:

• To develop a robust deblurring method that mitigates artifacts from inaccurate blur kernel estimation.
• To introduce a novel partial deconvolution model for improved image deblurring performance.
• To enhance the reliability of non-blind deconvolution techniques when dealing with imperfect blur kernels.

Main Methods:

• A partial map in the Fourier domain is proposed to model kernel estimation error.
• Reliable Fourier entries of the estimated blur kernel are detected to construct the partial map.
• Partial deconvolution is integrated with wavelet-based and learning-based models, utilizing an Expectation-Maximization (E-M) algorithm for alternative estimation of the partial map and latent image recovery.

Main Results:

• The proposed partial deconvolution model effectively suppresses artifacts stemming from inaccurate blur kernel estimation.
• Demonstrated significant improvements in deblurring quality on both synthetic and real-world blurry images.
• The E-M algorithm successfully estimates the partial map and recovers the latent sharp image.

Conclusions:

• The partial deconvolution model offers a robust solution for deblurring images with inaccurate blur kernels.
• This approach significantly reduces visual artifacts and enhances the overall quality of deblurred images.
• The method shows promise for practical applications in image restoration where precise kernel estimation is challenging.