| With the era of big data arrived,the need for high resolution image grows daily.Image super-resolution technology focus on how to reconstruct the high-resolution image from low-resolution image to fits more closely with human perception.Subject to the image degradation model,the high-resolution image reconstructed by the SISR based on machine learning theory will become worst when the assumptions are changed.With the analysis of SISR methods based on machine learning,the core algorithms of this paper is based on reconstruction as it can be executed more efficiently.The main ideas and innovations can be divided into two categories:1.With the reexamine of image degradation model,we find that the nature of image super-resolution problem can be regarded as a procedure of the removal of irrelevant pixel and the enhancement of image edge.The operation of removing irrelevant pixel can be realized by the image denoising algorithm,and the enhancement of image edge can be guaranteed by the theory of regularization.Based on these hypothesis,a new filtering-guided SISR algorithm is proposed in this paper.The experimental results show that FGSR can get a higher PSNR compared with other SISR methods not only under the condition of noiseless,but also in image degradation with much more noise.2.Through deep analysis of SISR methods based on machine learning,it demonstrates that the performance of SISR methods based on machine learning will be unstable as the bluer kernel in image degradation varies.In order to solve these problem,a fast non-blind single image super-resolution algorithm is proposed in this paper.Experiments show that on the one hand the high-resolution image can score a higher PSNR than other SISR methods,on the other hand the speed of FNSR meets the request of real-time and reliability.Compared with other SISR methods,the algorithms proposed here can get higher PSNR.But more importantly,it can run more efficiently and meet the requirements of engineering applications. |
PDF Full Text Request