Broad Spectrum Image Deblurring via an Adaptive Super-Network.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image blur varies significantly due to factors like camera shake and object motion.
• Current deep learning models struggle with diverse blur levels, leading to suboptimal deblurring.
• Specializing models for different blur levels while maintaining generalization is a key challenge.

Purpose of the Study:

• To propose Ada-Deblur, a super-network capable of deblurring images across a broad spectrum of blur levels.
• To dynamically adapt network architectures for specialized deblurring at various blur intensities.
• To achieve effective deblurring without requiring retraining on novel blurs.

Main Methods:

• Developed Ada-Deblur, a super-network architecture.
• Implemented dynamic network adaptation for flexible image processing at test time.
• Trained and evaluated the model on synthetic and realistic blurry images.

Main Results:

• Ada-Deblur outperforms strong baselines in reconstruction accuracy with minimal computational overhead.
• The method demonstrates effectiveness on both synthetic and realistic blurs.
• Significant performance gains, around 1 dB PSNR improvement, observed on unseen and strong blur levels.

Conclusions:

• Ada-Deblur successfully balances specialization and generalization across diverse blur levels.
• The dynamic architecture adaptation enables efficient and effective image deblurring.
• The proposed method offers a robust solution for handling a wide range of image blur scenarios.