Dual-Branch Discrimination Network Using Multiple Sparse Priors for Image Deblurring.

Area of Science:

• Computer Vision
• Image Restoration
• Deep Learning

Background:

• Blind image deblurring aims to recover sharp images from blurred observations.
• Traditional Convolutional Neural Networks (CNNs) struggle with real-world blur due to model-data incompatibility.
• Existing Generative Adversarial Networks (GANs) face training instability, hindering convergence.

Purpose of the Study:

• To propose a novel GAN architecture for effective blind image deblurring.
• To address the limitations of current deep learning models in handling complex image degradation.
• To improve the realism and detail in restored images.

Main Methods:

• Introduced a GAN with a dual-branch discriminator (DBSGAN).
• Incorporated multiple sparse priors into the discriminator's secondary branch.
• Balanced the generator-discriminator game for stable training.

Main Results:

• Demonstrated superior performance on synthetic and real-world blurry datasets.
• Achieved significant improvements in quantitative metrics and visual quality.
• Showcased a 1.7% average PSNR improvement on the GOPRO dataset compared to state-of-the-art methods.

Conclusions:

• The proposed DBSGAN effectively overcomes limitations in blind image deblurring.
• The dual-branch discriminator with sparse priors enhances training stability and restoration quality.
• The method offers state-of-the-art performance for realistic image deblurring.