Research On Image Restoration Algorithm Based On Deep Learning

A high-quality image is necessary to ensure the accurate transmission of information.Due to the operation of the photographer,the conditions of the scene,the weather conditions and other factors,most images captured by the imaging equipment will undergo unavoidable degradation phenomena,such as dynamic blurring,rain streaks and snow streaks.This phenomenon will greatly affect the quality of images and cause considerable losses to subsequent image information processing work,such as intelligent transportation,video surveillance,and other security-related visual applications.Therefore,restoring degraded images has high academic research and socio-economic value.The image restoration methods studied in this paper mainly include dynamic scene deblurring,single-image rain removal,and single-image snow removal.Among them,the dynamic scene blur is mainly because the image is affected by the relative motion between the camera and the object during the imaging process of the camera or other shooting equipment,and the blur phenomenon occurs.The rain streak attachment and snowflake attachment in the image are caused by bad weather,such as rainfall or snowfall during the imaging equipment shooting process.Unlike traditional image processing algorithms,image restoration research based on deep learning has attracted the attention of researchers and made breakthroughs.While analyzing various types of image degradation,this paper relies on various neural networks to study deblurring,rain removal,snow removal and other issues and proposes three effective image restoration algorithms based on deep learning.The main work of this study is as follows:(1)Aiming at the problem that the neural network in the image rain removal task requires a large amount of labelled data for training,and the problem of poor performance of the discriminator of the generative adversarial network in the field of image rain removal,a progressive recurrent attention neural network is proposed.The attention mechanism and the recurrent neural network are embedded into the multi-stage convolutional neural network,and a progressive recurrent loss function is used to better measure the effect of network training.At the same time,a progressive data filtering method is proposed,which can use a small amount of data to make the network achieve a better convergence effect.Finally,a novel algorithm using a pre-trained Transformer discriminator is proposed,which indirectly improves the generator’s rain removal performance by improving the discriminator’s classification performance in the generative adversarial network.(2)Aiming at the problem that the coarse-to-fine idea is not fully extended in the image deblurring task and the poor connection between multiple tasks,this paper proposes a multi-stage progressive encoder-decoder neural network based on the coarse-to-fine idea and extends the coarse-to-fine idea to the loss function level.Different from other progressive image deblurring networks,the deblurring network in this paper does not need to divide the input image into patches and can retain the global features of the image;finally,a novel image denoising method based on transfer learning is proposed.By using the end point of deblurring model training as the starting point of denoising model training,the model after deblurring training can be applied to single-image denoising.(3)Aiming at the problem of imperfect snow removal effect in single image snow removal task and the lack of long-distance dependence modelling ability of the pure convolutional neural network,a generative adversarial network integrating vision Transformer and convolutional neural network is proposed.By using the Transformer branch and the convolutional branch to extract and fuse features at the same time,the image generation ability of conditional generative adversarial network is used to improve the performance of the single image snow removal task.Experiments show that the network proposed in this paper has a more vital fitting ability than multiple methods that only contain convolutional neural networks,so the snow removal effect is better.