PRISM-deblur: Progressive refinement with intensity-guided spatio-temporal motion deblurring using RGB and event cameras

Abstract

Motion blur in dynamic scenes arises from both camera shaking and object motion, producing a mixture of global and local blur that is difficult to restore from RGB frames alone. While event cameras capture complementary motion information with high temporal resolution, effectively exploiting events remains challenging. Existing approaches often rely on shallow event encoders, unidirectional fusion that favors one modality, or coarse refinement strategies that cannot adapt to spatially variant blur. We propose PRISM, a progressive event-guided deblurring framework that introduces three dedicated modules. A Spatio-Temporal Event Processor (STEP) models temporal dynamics for robust event representation. A Bidirectional Cross-modal Attention Fusion (BiCAF) balances information exchange between image and event streams, addressing the limitations of unidirectional fusion in event-guided deblurring. A Motion-Intensity Guided Connection (MIGC) adaptively modulates refinement using motion cues, surpassing binary mask guidance. Extensive experiments on benchmark datasets demonstrate that PRISM achieves state-of-the-art performance, consistently outperforming prior methods in both quantitative metrics and perceptual quality. Qualitative comparisons further show that PRISM restores sharper motion details while preserving clean static regions, validating the effectiveness of our progressive design. Code and datasets are made publicly available at: github.com/Bingyu-Huang/PRISM.