Learning to Super-Resolve Blurry Images With Events.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Super-Resolution from a single motion Blurred image (SRB) is challenging due to combined low resolution and motion blur.
• Existing methods struggle with the ill-posed nature of SRB.

Purpose of the Study:

• To develop an Event-enhanced SRB (E-SRB) algorithm for generating high-resolution (HR) image sequences from single low-resolution (LR) motion-blurred images.
• To address the joint degradation of motion blurs, low spatial resolution, and event noise.

Main Methods:

• Formulated an event-enhanced degeneration model incorporating LR, motion blurs, and event noise.
• Developed an event-enhanced Sparse Learning Network (eSL-Net++) utilizing dual sparse learning for events and intensity frames.
• Proposed an event shuffle-and-merge scheme to extend single-frame SRB to sequence-frame SRB without retraining.

Main Results:

• The proposed eSL-Net++ significantly outperforms existing state-of-the-art methods on synthetic and real-world datasets.
• Demonstrated effective generation of sharp, clear HR image sequences from single LR motion-blurred images.
• Validated the algorithm's performance through extensive experimental results.

Conclusions:

• The E-SRB algorithm, powered by eSL-Net++, offers a robust solution for SRB problems.
• Event data integration effectively alleviates the challenges of motion blur and low resolution in image super-resolution.
• The proposed method advances the capabilities of single-image super-resolution techniques.