The Study On Pixel-level Non-local Self-similarity And Cycle-consistency For Image Restoration

Objective: Image restoration(such as image super-resolution,image denoising,image de-artifacts,image enhancement,etc.)is a long-standing low-level problem in computer vision,which aims to recover to high-quality images from low-quality images(such as down sampled,noisy,and compressed images,etc.).Currently,image restoration is still an important problem in computer vision.Image restoration not only helps obtain a more pleasing visual experience but also has profound auxiliary significance for some specific downstream tasks,such as image segmentation,detection,recognition,etc.Early image restoration methods based on image blocks were often implemented by introducing predefined block-level self-similarity priors and wavelet transforms with fixed parameters.Although some results have been obtained,there is still some distance from the practical application.Moreover,in recent years,due to the good performance of convolutional neural networks(CNN)in learning generalizable image priors from large-scale data,these models have been widely used in image restoration and related tasks.Although CNN-based methods have achieved more robust performance than traditional methods,the current discriminativebased CNN model still needs to be improved for tasks that lack training data.There is still much room for improvement in restoration performance.In order to achieve a more stable and efficient image restoration performance,this paper focuses on the rapid de-artifacts reconstruction of magnetic resonance images,optical coherence tomography image denoising,and image super-resolution based on pixel-level non-local methods and deep learning technology,the algorithm for the above tasks and the model made a more systematic and in-depth exploration and proposed three more robust image restoration methods.Methods and results: Aiming at the problem of low-resolution images,this paper proposes a new super-resolution model based on semi-cycle generative adversarial networks;because of the slow speed of MRI and the introduction of more artifacts after under-sampling accelerated imaging,a new model is proposed.An iterative refinement reconstruction model for compressed sensing magnetic resonance images based on pixel-level non-local selfsimilarity prior;Aiming at the problem that optical coherence tomography images contain strong speckle noise,a denoising method based on pixel-level non-local self-similarity prior is proposed.(1)This paper firstly proposes a denoising method PNOCT based on pixel-level nonlocal self-similarity for optical coherence tomography image denoising.The innovation is that a novel pixel-level self-similarity prior,by exploring the pixel-level self-similarity in the image,a better denoising effect can be achieved,and the high-frequency details in the image can be preserved more completely.(2)This paper also proposes a pixel-level self-similarity based model PNCS for the accelerated reconstruction of magnetic resonance images.The iterative refinement model proposed can effectively remove the artefacts introduced in the process of reconstruction in the sampling and preserve details.(3)This paper further proposes a semi-cycled Generative Adversarial Network model SCGAN for image super-resolution.The innovation is that,compared with the previously proposed fully-cycled cycle consistency generative adversarial networks,the proposed SCGAN establishes two independent sub-networks for learning the degradation process of images to learn the degradation process of real-world high-resolution images and synthetic high-resolution images,respectively,and inputs the two degradation results into the same restoration sub-network to establish a low-resolution sub-network.Mapping of highresolution images to high-resolution image domains to improve the generalization ability of the model to better cope with low-resolution images of different types and degrees of degradation processes.Conclusion: This paper proposes three innovative natural and medical image restoration methods,which respectively realize undersampling accelerated reconstruction of MRI images,denoising of optical coherence tomography images,and super-resolution of face images.Compared with the existing methods,PNOCT can achieve more advanced optical coherence tomography image denoising performance while alleviating highfrequency loss of information;PNCS can better remove artifacts in undersampled MRI images and better preserve image details;SCGAN can obtain more realistic image superresolution results.The above methods and models all have a robust auxiliary effect on downstream tasks and have high application and follow-up research value.