Motion-Adaptive Deblurring of Fast Moving Objects

Fast Moving objects (FMOs) captured by cameras often exhibit significant blur and artifacts. General deblurring algorithms face challenges in accurately restoring the appearance and intricate details of severely blurry regions within an image. To address these limitations and facilitate single image deblurring of FMOs, we propose a novel Motion-Adaptive FMOs image deblurring Network named MANet. Specifically, to effectively leverage the motion information of FMOs and guide the network to pay attention to blurry regions, we introduce three key modules including a motion-adaptive spatial channel attention module, a blur region identification module, and an adaptive dual branch module. The motion-adaptive spatial channel attention module incorporates motion cues and propagates them to subsequent restoration modules at each scale. Meanwhile, the blur region identification module precisely locates blurry regions for targeted processing. Furthermore, considering that blurry regions and clear regions exhibit distinct feature representations, the adaptive dual branch module is introduced to independently extract features from these two types of regions and subsequently fuse them rationally. Extensive evaluations on benchmark datasets demonstrate the proposed MANet performs favorably against general deblurring methods and existing FMO-specific deblurring approaches.