Zhou Shangchen ; Li Chongyi ; Loy Chen Change

Abstract

Night photography typically suffers from both low light and blurring issuesdue to the dim environment and the common use of long exposure. While existinglight enhancement and deblurring methods could deal with each problemindividually, a cascade of such methods cannot work harmoniously to cope wellwith joint degradation of visibility and textures. Training an end-to-endnetwork is also infeasible as no paired data is available to characterize thecoexistence of low light and blurs. We address the problem by introducing anovel data synthesis pipeline that models realistic low-light blurringdegradations. With the pipeline, we present the first large-scale dataset forjoint low-light enhancement and deblurring. The dataset, LOL-Blur, contains12,000 low-blur/normal-sharp pairs with diverse darkness and motion blurs indifferent scenarios. We further present an effective network, named LEDNet, toperform joint low-light enhancement and deblurring. Our network is unique as itis specially designed to consider the synergy between the two inter-connectedtasks. Both the proposed dataset and network provide a foundation for thischallenging joint task. Extensive experiments demonstrate the effectiveness ofour method on both synthetic and real-world datasets.