Deep Attention-based Lightweight Network For Aerial Image Deblurring

Aiming to challenge that current image deblurring methods fail to deal with complex scenes well and are computationally time-consuming, we present an attention-based lightweight network (MALNET) that restores sharp images in a pyramid learning pattern. The model introduces the lightweight chained residual unit as the core building block to construct the multi-path refinement network and embeds the simple yet efficient attention unit to obtain rich contextual information. Moreover, we explore multiple training strategies such as multi-scale joint loss and parameter sharing, making MALNET more flexible and compatible in various application scenarios. The experimental results sampled from the DOTA and GoPro datasets demonstrate that MALNET can enjoy 5 times faster speed and 7 times fewer bulk than those state-of-the-art approaches while still providing a competitive performance both quantitatively and qualitatively.