Two-stage single image Deblurring network based on deblur kernel estimation
Ying Cheng Lu,Tzu Pu Liu,Chang Hong Lin
DOI: https://doi.org/10.1007/s11042-022-14116-z
IF: 2.577
2022-11-08
Multimedia Tools and Applications
Abstract:Image deblurring for dynamic scenes is a serious challenge in computer vision. Motion blur is caused by camera shaking or object movement during the exposure time. Many photos cannot be reproduced at the moment they were taken, its contents cannot be restored if motion blur occurs. In this article, we proposed a deblurring system that uses a two-stage convolutional neural network (CNN) to achieve image deblurring through a joint learning strategy. The first-stage network predicts the deblur kernel of each pixel and pre-deblurs the input image, and then the second-stage network directly predicts clear images based on U-Net architecture. In the first-stage network, the deblur kernel uses the surrounding information to restore the centre pixel, which can effectively remove the tiny motion blur. To additionally deal with large motion blur, we extend the second-stage network is used to compensate for the limited receptive field of the first-stage deblurring kernel. We evaluate the proposed method on benchmark blur datasets. Experimental results show that the proposed method can produce better results than state-of-the-art methods, both quantitatively and qualitatively. The proposed method can achieve the best PSNR at 32.59db, 27.21db and 31.96db for the GOPRO, Köhler, and Su datasets, respectively.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering