| With the continuous development of super-resolution technology in recent years,super-resolution has been widely applied in various real-world scenarios,such as remote sensing imaging,image enhancement,photo restoration,and data compression.The degradation model is an important component of superresolution technology,as it determines the scenarios targeted by super-resolution techniques and the types and levels of degradation to be restored in the images.As super-resolution is increasingly applied in real-world scenarios,traditional degradation models gradually become insufficient to cover the diverse degradation scenarios in the real world.This has led to the emergence of various blind super-resolution methods,which are super-resolution methods designed for unknown image degradation.The degradation model plays a crucial role in blind super-resolution models,and designing a reasonable degradation model to handle real-world degradations is a key aspect of blind super-resolution.To address the blind super-resolution in real-world image scenarios,this study proposes a more practical degradation model.The model integrates various novel degradation methods and introduces a novel random and shuffled degradation strategy,highlighting the randomness of degradation types and lengths,corresponding to the complexity of the real world.In terms of degradation types,the model innovatively incorporates traditional degradation types such as multi-step downsampling and rotational degradation,which were not included in previous degradation models.Regarding the lengths of degradation sequences,the model employs a shuffling and random deletion strategy to generate images with different levels of degradation.Experimental results demonstrate that the proposed random degradation model outperforms traditional degradation models in terms of PSNR/SSIM metrics on the DIV2KRK dataset.This indicates that the degradation model effectively expands the degradation space,thereby better simulating real-world degradation scenarios.Based on the proposed random degradation model,this study proposes a baseline super-resolution network(RDNet)and further extends it to a superresolution network(RDGAN)based on the random degradation model,utilizing PatchGAN and VGG19 feature extractor.Additionally,a new evaluation dataset called DIV3T is created,and comprehensive experimental comparisons are conducted on this dataset to evaluate the proposed RDNet,RDGAN,and similar previous works.The experimental results show that the proposed superresolution method outperforms other blind super-resolution methods in terms of PSNR/SSIM metrics.Moreover,in terms of subjective visual perception,the proposed super-resolution network enhances image details and recovers clean and sharp images from low-resolution inputs,demonstrating superior performance in the task of super-resolution reconstruction compared to previous methods.In conclusion,the proposed approach effectively improves the quality of super-resolution in data close to real-world scenarios and provides valuable insights for blind super-resolution applications in real-world settings. |
