Sub-Pixel Generative Adversarial Network And Its Application In Image Deblurring

Image deblurring is a subproblem in image restoration,it is a challenging task in computer vision and image processing.Most of the early traditional deblurring methods(such as inverse filtering,Lucy-Richardson algorithm,Wiener filtering,Tikhonov regularization,etc.)perform non-blind deconvolution operations on the blurry process to obtain the estimation of clear images under the assumption that the blur kernel is known.But these methods have strong assumptions,and there are different degrees of disadvantages.For example,inverse filtering algorithms are more sensitive to the interference in the noise,it is difficult to restore the clear image from the blurred image with noise.The Tikhonov regularization method is poor in restoring the details in the complex structure images.Since Goodfellow et al.proposed the model of generating adverse network(GAN)in 2014,it has shown strong learning ability in the area of image processing(such as image generation,image inpainting,super resolution,etc.),which has won the favor of many scholars.Compared with many traditional image restoration methods,GAN is helpful to improve the quality of reconstructed images.Since then,many researchers have proposed a variety of deblurring models based on this framework and achieved excellent results.A representative of these GAN-based models is the Deblur GAN model proposed by Kupyn et al.,which achieves the-stateof-the-art(SOTA)in terms of both structural similarity measures and visual appearance compared to other deblurring methods,and it achieves perceptually convincing results.In the Deblur GAN,however,due to the transposed convolution(or deconvolution)operation in the generator,there are many unnecessary 0 fill elements during the upsampling of the image,resulting in checkboard artifacts in the restored image.Therefore,subpixel convolution is used in this paper to replace transposed convolution in Deblur GAN.In this way,there is no need to add meaningless 0 elements in the upsampling process,thus it is able to eliminate the checkboard artifacts produced by transposed convolution to some extent,and it can effectively reduce the training cost.The main work and research contents of this thesis are as follows:1.The research status of image deblurring is described.The model framework,advantages and disadvantages of various deblurring methods based on generative adversarial network are analyzed in details.2.An sub-pixel generative adversarial network based image deblurring method is proposed.We combined the sub-pixel convolutional neural network with Deblur GAN,replacing the transposed convolution in the original Deblur GAN with sub-pixel convolution,and proposed an architecture called Sub-Pixel Generative Adversarial Network(Sub-pixel GAN),which can effectively mitigate the checkboard artifacts contained in the restored image.In this thesis,the effectiveness of the proposed model is verified by experimental analysis.Compared with Deblur GAN and other related methods,the subpixel GAN can eliminate the checkerboard artifacts in blurred images to a certain extent,and achieves good results on several commonly used benchmark data sets.