Motion Aware Double Attention Network for Dynamic Scene Deblurring

Motion deblurring in dynamic scenes is a challengingtask when the blurring is caused by one or a combination ofvarious reasons such as moving objects, camera movement,etc. Since event cameras can detect changes in intensitywith a low latency, necessary motion information is inherently captured in event data, which could be quite usefulfor deblurring standard camera images. The degradationintensity does not show homogeneity across an image dueto factors like object depth, speed, etc. We propose a twobranch network structure, Motion Aware Double AttentionNetwork (MADANet), that pays special attention to areaswith high blur. As part of the network, event data is firstused by the high blur region segmentation module that creates a probability-like score for areas exhibiting high relative motion to the camera. Then, the event data is alsoinjected to feature maps in the main body, where there isa second attention mechanism available for each branch.The effective usage of event data and two-level attentionmechanisms makes the network very compact. During theexperiment, it was shown that the proposed network couldachieve state-of-the-art performance not only on the benchmark dataset from GoPro, but also on two newly collecteddatasets, one of which contains real event data