Deep Dynamic Scene Deblurring for Unconstrained Dual-Lens Cameras.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Dual-lens (DL) cameras capture depth information for advanced vision applications.
• Unconstrained settings in DL cameras are common but lead to motion blur challenges.
• Existing deblurring methods fail with unconstrained DL cameras, causing view inconsistencies and disrupting disparities.

Purpose of the Study:

• To address dynamic scene deblurring for unconstrained dual-lens cameras.
• To overcome view-inconsistency and preserve scene-consistent disparities in deblurred images.
• To develop deep learning techniques for space-variant and image-dependent motion blur.

Main Methods:

• Developed a Coherent Fusion Module to resolve view-inconsistency in deblurring architectures.
• Introduced a memory-efficient Adaptive Scale-space Approach to handle varying image scales without parameter increase.
• Proposed a module to tackle the space-variant and image-dependent nature of dynamic scene blur.

Main Results:

• The proposed Coherent Fusion Module effectively reduces view-inconsistency.
• The Adaptive Scale-space Approach maintains scene-consistent disparities efficiently.
• The new module successfully addresses complex dynamic motion blur characteristics.
• Experimental results demonstrate substantial practical merit of the proposed techniques.

Conclusions:

• The developed deep learning framework provides a robust solution for dynamic scene deblurring in unconstrained dual-lens cameras.
• The novel modules significantly improve image quality and disparity consistency in challenging motion scenarios.
• This work advances the capabilities of dual-lens cameras in capturing clear images under motion.