Research And Implementation Of Image Super-Resolution Reconstruction Method Based On Machine Learning

Super-resolution reconstruction technology aims to reconstruct a high-resolution image from a degraded low-resolution image.In recent years,with the progress of computer hardware technology,super-resolution reconstruction technology based on convolutional neural networks has also made great progress,but the existing super-resolution technology still has certain problems in practical applications.From the three perspectives of super-resolution technology,this article proposes corresponding solutions to the related problems of current super-resolution technology.The main work is as follows:(1)This paper proposes a super-resolution network for general static degradation model.Existing super-resolution reconstruction methods based on convolutional neural networks simplify the degradation model of low-resolution images.It is only assumed that low-resolution images are downsampled by bicubic interpolation of high-resolution images.This makes it difficult for the existing algorithms to have a good effect on low-resolution images observed in practical applications.Under the more generalized image degradation model,this solution uses the degradation factor prediction of the DFPN module and the high-resolution reconstruction network of the HRN module to generate the final super-resolution image.The PSNR and SSIM scores of multiple sets of comparative experiments prove the superiority of the SR-GSD algorithm for more general low-resolution image reconstruction effects.(2)This paper proposes a lightweight neural network architecture search algorithm SRNAS for super-resolution.In recent years,researchers often use deeper convolutional neural networks to fit the mapping function from low-resolution images to high-resolution images.Although the algorithm's reconstruction performance is improved,it also increases the computational complexity.This solution applies NAS technology to image super-resolution,and uses a well-designed search space and search strategy to generate neural network models.Experiments show that the final searched SRNAS network model can greatly reduce the computational complexity and weight parameters of the algorithm,while maintaining the reconstruction accuracy almost unchanged.(3)This paper builds an easy-to-use image super-resolution repair platform.Although the current super-resolution reconstruction technology has achieved certain development,for ordinary users,the current super-resolution reconstruction technology has a high threshold to use,and users need to learn a lot of relevant knowledge and technology to reconstruct existing low-resolution images.In order to further reduce the application difficulty of super-resolution reconstruction technology,this paper builds an image super-resolution repair platform based on flask microservice framework and pytorch deep learning framework,it is convenient for users to choose the appropriate super-resolution reconstruction method and scaling scale for image reconstruction according to their needs.