Nonblind Image Deblurring via Deep Learning in Complex Field.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Nonblind image deblurring is a challenging inverse problem in imaging.
• Recovering clear images from blurry ones requires robust noise suppression.
• Existing methods often struggle with unknown noise distributions during deblurring.

Purpose of the Study:

• To develop a superior deep-learning-based method for nonblind image deblurring.
• To introduce a novel image prior in the Gabor domain for improved noise handling.
• To leverage complex-valued representations for enhanced denoising capabilities.

Main Methods:

• A convolutional neural network (CNN)-based image prior was defined in the Gabor domain.
• Complex-valued (CV) representations were employed in intermediate processing for better denoising.
• A CV CNN was developed for improved generalization to unknown noise types.
• The Gabor-domain CV CNN prior was integrated with an unrolling scheme for deblurring.

Main Results:

• The proposed approach demonstrated superior performance in nonblind image deblurring.
• The Gabor transform's space-frequency resolution and orientation selectivity were exploited.
• Complex-valued CNNs showed better generalization for handling unknown and varying noise.
• Experimental results confirmed the effectiveness over state-of-the-art methods.

Conclusions:

• The novel Gabor-domain complex-valued CNN prior offers significant improvements in nonblind image deblurring.
• This deep learning approach effectively addresses noise magnification challenges.
• The method provides a robust and high-performing solution for image restoration tasks.