EUMD: Event-based Unrolled Motion Deblurring via Learnable Prior Optimization

초록

Motion blur caused by finite exposure times remains a significant challenge for conventional cameras, degrading image quality and downstream computer vision tasks. Event cameras, which asynchronously report brightness changes with microsecond latency, capture high-frequency temporal dynamics that are lost in standard frame-based sensing. In this work, we propose Event-based Unrolled Motion Deblurring (EUMD), a hybrid, optimization-driven framework that tightly couples the continuous-time blur formation model with asynchronous event dynamics. We formulate deblurring as a maximum a posteriori (MAP) problem built upon a novel, event-conditioned diagonal forward operator. Unlike traditional methods that rely on fixed, handcrafted priors such as image gradients, our framework adopts a flexible design where the prior is modeled via a learnable deep network. To solve the resulting ill-posed inverse problem, we employ Half-Quadratic Splitting (HQS) and unroll its iterations into a K-stage deep network. The proximal operator is implemented using a Transformerenhanced U-Net, which captures both hierarchical local structures (via a CNN backbone) and long-range non-local dependencies (via a Transformer bottleneck), and is adaptively conditioned on event information. This unrolled architecture is interpretable, efficient, and fully end-to-end trainable, allowing both the optimization parameters and the deep prior to be learned jointly. Extensive experiments on synthetic (GoPro) and real-world (EVRB) benchmarks demonstrate that EUMD achieves state-of-the-art performance, outperforming existing event-guided approaches. © 2026 IEEE.

키워드