ISBN: 978-981-18-5852-9 DOI: 10.18178/wcse.2022.04.057
Depth Separable Convolution Lightweight Blind Motion Deblurring Algorithm
Abstract— Aiming at the motion blur problem caused by handheld camera shake and object movement in the process of photo acquisition, an image deblur algorithm based on depth separable convolution lightweight generative adversarial network was proposed. Firstly, a depth separable convolution residual block with selfattention mechanism is introduced in the generator. Secondly, Transpose convolution module is replaced by the Pixel-Shuffle up-sampling module. Finally, The combination of perceptual loss and relativistic generative adversarial loss can effectively alleviates the mode collapse problem of traditional GAN and improves the stability of model training. Experimental results show that compared with Deblur-GAN the number of parameters of the proposed algorithm is significantly reduced, the deblurring time of a single image is reduced by 50%, the deblurring image is clearer in subjective vision, and the objective evaluation indexes of image quality such as peak signal-to-noise ratio and structural similarity are also significantly improved.
Index Terms— image deblurring, Generative adversarial nets, self-attention, deep learning, depth separable convolution, relativistic average discriminator.
Chi Li
School of Computer Science and Technology, Xi'an University of Posts and Telecommunications, China; Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi'an University of Posts and Telecommunications, China
Weiwei Kong
School of Computer Science and Technology, Xi'an University of Posts and Telecommunications, China; Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi'an University of Posts and Telecommunications, China
Jiawei Xue
School of Computer Science and Technology, Xi'an University of Posts and Telecommunications, China; Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi'an University of Posts and Telecommunications, China
Ze Wang
School of Computer Science and Technology, Xi'an University of Posts and Telecommunications, China; Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi'an University of Posts and Telecommunications, China
Cite: Chi Li, Weiwei Kong, Jiawei Xue, Ze Wang, " Depth Separable Convolution Lightweight Blind Motion Deblurring Algorithm " WCSE 2022 Spring Event: 2022 9th International Conference on Industrial Engineering and Applications, pp. 484-493, Sanya, China, April 15-18, 2022.