Research On Image Restoration Technology Based On Deep Learning

The application of deep learning to computer vision has been a popular research direction in recent years,and many model algorithms of practical significance have emerged and are widely used in various industries.Image restoration based on deep learning is one of the important ones.As the research progresses,the usefulness of restoration networks continues to increase,from restoring fixed rectangular-shaped regions to arbitrarily shaped regions.In addition,the restoration results are becoming more and more realistic,and it is no longer only required to restore results similar to the original image,but also extended to generate brand new images with consistent style and cover more restoration scenarios.However,there are still many problems and room for improvement.For example,restoration for larger size edge areas is prone to blurring and incompleteness,while the complementary content tends to be monotonous and repetitive,and the consistency with the original image is low.In addition,the restoration for larger irregular regions is often semantically ambiguous,prone to chromatic aberrations and distortions,and the texture details are not clear enough.In order to solve the above problems,this paper proposes two image restoration models based on deep learning theory and existing restoration models,and the main research contents are as follows.First,for the problem of content blurring in the restoration of large－sized edge regions,the paper proposes a multi-scale feature-based image restoration model,which ensures image consistency while enriching detailed textures and improving the clarity of the restored region.A two-stage generative adversarial network structure is used,with the first stage restoring the edge information of the image and generating a black-and-white edge information image,and the second stage using the edge image generated in the first stage to complete the color image and generate the final result.A parallel dilation convolution structure is added to the edge restoration network,while a multi-scale dilation convolution fusion structure is added to the image complementation network,and feature content loss is added to expand the perceptual field and enrich the scale of feature extraction at the same time.Through experiments,the model proposed in this paper has better restoration results on larger edge regions,which are verified by commonly used evaluation metrics such as PSNR and SSIM.Second,to address the problems of structural errors and ambiguity in the restoration of arbitrarily shaped regions,the paper proposes a two-stage recursive feature inference-based model for restoration of arbitrarily shaped regions of images,which improves the accuracy of the restoration result structure,ensures the correct semantic information,and also makes the results of restoration regions with larger sizes clearer.In this paper,the proposed model uses different repair modules recursively in two stages.The first stage uses the dual-scale merged repair module for initial repair to obtain a coarser repair result,while using semantic feature loss to ensure the accuracy of the semantic information of the reconstruction result.In the second stage,the recursive restoration is performed using the consistent attention restoration module to make full use of the information in the effective part of the image and enrich the texture details of the reconstruction results.For each restoration in the second stage,the part of the restoration obtained by this module is retained using the continuously updated mask,and the final result is obtained by the feature fusion module.Finally,the effectiveness of the model is verified by the commonly used evaluation metrics such as PSNR and SSIM.Finally,an image restoration verification and evaluation platform is built based on the two restoration models proposed in this paper,and the two models proposed in this paper are deployed into the platform to facilitate the verification and evaluation of the algorithmic models in this paper.The platform is based on B/S architecture,providing the function of uploading images for restoration or extension to the outside world,and the subjective effect of the restored images can be observed and evaluated directly after the completion of processing,or downloaded for verification in other ways.In addition,the image extension module based on the edge restoration model provides a manual guidance function to draw black and white edge images online to directly guide the final color generation results.