Research And Implementation On Image Super-Resolution Based On Generative Adversarial Networks

The process of converting low-resolution images into high-resolution images is called image super-resolution generation,and is a hot research topic in the field of images in recent years.The super-resolution generation of images has important needs and functions in the fields of image transmission,aerial remote sensing,medical analysis,surveillance and so on.How to improve the quality and the speed of the image generation to achieve the level of practical application is an urgent problem to be solved.In this thesis,the super-resolution algorithm based on generative adversarial network is improved,the image quality is improved,the model is compressed to reduce the running time,and super-resolution technology is used to complete the design and implementation of the image enhancement system.The main work of this thesis is as follows:(1)For the problem that the super-resolution algorithm based on generative adversarial network has confusing image details and low credibility,the loss function based on edge detection and the generator based on DenseNet model are proposed.The method uses edge detection to form the contour boundary and texture features of the object,which limits the generation of unnatural texture details.The DenseNet network model is used in the generator to increase the reuse of image features,reduce the complexity of the model and improve the image quality.After experimenting on the SET5?SET 14 and BSD 100 datasets,compared with the SRGAN algorithm,the image details generated by this method are clearer and more efficient,the PSNR and SSIM indicators are increased by 3.9%and 3.5%,and the parameters of the generator model is reduced by 90%.(2)For the problem that the super-resolution algorithm based on the generative adversarial network has poor real-time image performance,a new compression method,namely iterative generative adversarial network compression model,is proposed.The model iteratively simplifies the network model by changing the input model of the generative adversarial network,reducing the time-consuming of the network.Experiments show that,verified by handwritten dataset experiment,the runtime of super-resolution generation using model compression,is up to 17 times faster than before.(3)Design and implement a super-resolution image enhancement system based on the generative adversarial network.The system realizes the modules of image compression transmission,client super-resolution generation and online training by building a graphic website platform.It reduces the bandwidth consumption of the website by 75%,and provides model training and fine-tune services for image super-resolution generation of professional dataset.