Efficient multi-stage network with pixel-wise degradation prediction for real-time motion deblurring

Motion deblurring is an indispensable task in perceptual systems because motion blur can seriously influence the visual effect and the quality of subsequent perception tasks. In practical application, efficiency and effect of motion deblurring are both important. However, among the existing methods, there is no design that can meet the real-time and visual quality requirements at the same time. Thus, an efficient motion deblurring network is proposed by leveraging multi-stage pixel-wise degradation prediction. Specifically, some lightweight modules are designed to accelerate processing while attention and multi-scale mechanism are introduced to maintain quality. In addition, a pixel-wise degradation prediction module and a spatial-channel compensation module are further employed to improve the deblurring quality, such as the distortion of moving objects in the restoration images. Extensive experimental results show that the proposed network can achieve the same PSNR level as the SOTA lightweight deblurring methods and is far faster (5.3 times for DMPHN, 6.7 times for IFI-RNN). Therefore, the proposed design achieves a balance between quality and speed compared with the existing methods.