Image Restoration Based On Attention Mechanism Generative Adversarial Networks

In the process of imaging,the images captured often encounter atmospheric pollution,reflection,transmission,and other conditions,resulting in image degradation problems.Image degradation will not only affect the visual quality of the image,but also change the vital information contained in the image,and then affect the use of some vision systems and the research of some subsequent computer vision algorithms.Therefore,it is important to study high-quality digital image restoration methods.Traditional image restoration methods have many disadvantages,such as constructing a complex degradation model,relying on prior knowledge,and so on.In recent years,the image restoration methods based on convolutional neural networks have adopted a data-driven method to learn features,which can realize the linear or nonlinear feature mapping between tasks,and overcome the limitation of specific prior knowledge.This kind of method recovers the degraded image by obtaining the global features of the image and has achieved success in the field of image processing.Compared with the methods based on the convolutional neural network,the image restoration methods based on the generative adversarial network can restore the original image with higher resolution and have achieved a series of excellent results in various image processing tasks,but its poor universality and the need for paired datasets to train the network to affect the further application of the generative adversarial network in the field of image restoration tasks.Because of the above problems,this thesis uses the combination of generative adversarial network and visual attention mechanism to solve the problem of image restoration with the idea of generation.This dissertation mainly studies the image restoration of the generative adversarial network in two specific situations: image dehazing and bleedthrough image restoration,through theoretical analysis and experimental verification,this thesis makes a beneficial exploration of the improvement and application of generative adversarial network.The main research contents of this dissertation are as follows:(1)Affected by the bad weather of haze,the use of many practical applications such as video surveillance and automatic driving are vulnerable and are limited,and computer vision tasks such as object positioning and detection are difficult to complete.Existing image dehazing methods often rely on the prior information in the atmospheric scattering model or supervised learning solution based on paired images,an image dehazing method based on the attention mechanism of the cycle-consistent generative adversarial network is proposed,which uses the constraints transfer learning ability and cyclic structure of the cycle-consistent generative adversarial network to complete the processing of unpaired images for unsupervised dehazing task.Considering the complexity of haze distribution in actual imaging,combined with human visual characteristics,the channel attention mechanism,and domain attention mechanism are integrated into the network to deal with different features and regions unevenly.The experimental results show that the algorithm can avoid the color distortion and other problems in processing the synthetic image datasets and real hazy image datasets,and have a good dehazing effect.(2)When a document is scanned as a digital image,the content on the back will usually be transmitted to the content on the front,which affects the normal reading and use of the document scanned image.Most existing single bleed-through image methods require learning supervised models from a large set of paired synthetic training datasets,which significantly limits their scalability and practicality.In this thesis,an image restoration method based on self-supervised constraints is proposed by using the cycle-consistent generative adversarial network to achieve unsupervised single image bleed-through removal of unpaired datasets.The network input image first passes through the feature enhancement module to extract features to enhance the texture details of the input image.A self-learning module is added to the network to introduce self-supervised constraints from the feature information of unpaired bleed-through images and non-bleed-through images,extract effective features from shallow layers to deep layers,improve the restoration quality of texture information,and retain the content and details of the front image as much as possible.Experimental results show that the proposed method achieves good bleed-through image restoration results on the synthetic image datasets,public datasets,and real bleed-through image datasets.