Research On Image Denoising And Super Resolution Reconstruction Technology

Image denoising and super-resolution reconstruction are important technologies in the field of image processing,which have important applications in many fields such as national defense,aerospace,medical and so on.Image denoising is the basis of improving the performance of advanced computer vision tasks such as classification,segmentation and object recognition,and it is also an important part of solving various computer vision problems.Classical image denoising and super-resolution reconstruction methods are mainly filter based methods and interpolation based methods.These methods generally have the problem of poor universality for different scenes.Existing research results show that deep learning for image denoising and super-resolution reconstruction usually has better performance than traditional machine learning methods.However,most of the existing researches treat image denoising and super-resolution reconstruction as two different topics.Therefore,the first work of this paper integrates these two technologies in order to improve the processing efficiency and obtain better processing quality.In addition,the second work of this paper focuses on the non-local features in image denoising,and uses non-local similarity to obtain better image denoising performance.Specifically,this paper mainly includes the following two works:(1)In the first work,aiming at the problem of denoising and high-resolution reconstruction of noisy images,a novel multi-scale fusion adversarial network model based on generative adversarial network is proposed to integrate image denoising and super-resolution reconstruction tasks,so as to restore the image quality affected by noise.In this way,the combination of image denoising and image super-resolution reconstruction simplifies the process of up-sampling and down-sampling in the process of model learning and analysis,avoids repeated input-output image operations,and effectively improves the efficiency of the entire image analysis process.Another important improvement of the proposed model is the multi-scale feature fusion technique,which uses multiple convolution kernels of different sizes to expand the receptive field in parallel,and in conjunction with the fusion of a new loss metric for perceptual loss and adversarial loss to optimize network learning.Experimental studies show that the proposed model benefits from the multi-feature perception ability of multiple receptive fields,and can achieve better high-resolution image reconstruction performance than the existing methods.Meanwhile,ablation experiments also verify the effectiveness of each of the new loss measures designed by us.(2)In the second work,a novel image denoising method based on graph convolutional neural network is proposed to overcome the defect that existing image denoising algorithms cannot obtain non-local similarity.The proposed method creates neurons with non-local receptive fields based on graph convolution operation.Graph convolution is a generalization of traditional convolution operation,which is used to deal with data with irregular structure.Graph convolution operations generalize classical convolution operations to arbitrary graph structures.In our proposed method,the graph structure is dynamically constructed according to the similarity between the features of hidden layers in the network,so as to express the self-similarity by using the powerful representation learning ability of graph convolutional network.More specifically,we use graph convolutional network layer to define the neighborhood of each pixel in a flexible way.Using this method,we can extract not only the features that depend on the spatial neighbors,but also the features that depend on the spatial distance pixels,which can show the latent feature correlation in the hidden space.Experimental results show that the proposed model benefits from the feature sensing ability of non-local graph convolution operation,and can achieve better image denoising performance than the existing classical image denoising methods.