Research And Application Of Image Deblurring Method Based On Deep Learnin

Images acquired by imaging devices are blurry due to factors such as camera shake,outof-focus or object motion,which seriously affect the demand for high quality of clear images.Therefore,the image deblurring problem has been the focus of research in image restoration techniques.Traditional image deblurring methods based on blur kernel estimation are limited in scope and vulnerable to unknown noise,and there is still a ringing effect in the recovered image,so the deblurring effect is poor.While image deblurring methods based on deep learning suffer from the loss of details in the restored images.Based on this background,three image deblurring methods based on deep learning are further studied in this paper.The main work and innovation points are as follows:(1)Aiming at the problem of ringing effect in traditional image deblurring methods,an end-to-end image deblurring method based on generative adversarial network is proposed.The feature enhancement module is added into the feature pyramid network for improvement,and the improved feature pyramid network is used as a generator to generate restored images.The final deblurring image is obtained through the adversarial training of the discriminator.Experiments show that this method can effectively remove image blur and reduce the ringing effect in image restoration,and the deblurring performance is better than those traditional image deblurring methods.(2)Aiming at the problem of missing details in restored images by existing image deblurring methods based on deep learning,an image deblurring method of image patches based on multi-scale convolutional neural network is constructed.The method uses a multi-scale convolutional neural network to optimize the image from fine to coarse,aggregating multiple image patch features and adding channel enhancement attention modules into the encoding network to improve the deblurring effect.Experiments show that this method can significantly reduce the ringing effect in image restoration and restore image details compared with existing image deblurring methods based on deep learning.(3)Aiming at the problem that general deep learning network models cannot accurately establish the formation process of image blur and rely on training datasets,an image deblurring method with combined deep priors is proposed based on image prior information.Deep priors of the clear latent image and the blur kernel are modeled respectively by the self-encoding-selfdecoding network and the fully connected network,and the final clear image is obtained through joint optimization.The deblurring model is also improved by adding TV regularization and impact filtering,so as to improve the noise-resistance performance of the method and enhance the details of the restored image.Experiments show that this method can effectively extract the prior information from blurry images and recover image details without relying on a large number of training datasets,with less obvious ringing effect and higher adaptivity.