Event-Based Motion Deblurring with Dual Channel Attention

Abstract

The event camera with high temporal resolution has shown promising advancements in motion deblurring tasks. However, due to the imperfect structure of network and the neglect of mutual compensation between events and images, existing event-based methods struggle to undertake effective feature extraction, leading to limited deblurring performance. To this end, we proposed a novel Dual Channel Attention Network (DCA-Net) by designing a Dual Channel Attention Block (DCAB), which leverages the benefits of multi-head attention and channel attention to sufficiently extract features from different dimensions and facilitate interaction of dual features. Thus the network can efficiently extract features from both events and images, enabling simultaneous enhancement of blurry features and suppression of event noise. Additionally, a decoder component based on Multi-Layer Perceptrons (MLP) is integrated into the network to facilitate continuous-time reconstruction. Extensive experimental results on synthetic, semi-synthetic, and real-world datasets demonstrate that our DCA-Net outperforms state-of-the-art motion deblurring methods.