Single Image Super-Resolution Using Feature Adaptive Learning And Global Structure Sparsity

Image super-resolution refers to recovering missing information from the low-resolution image to obtain the corresponding high-resolution image using available prior knowledge,which is a hot research direction in the field of digital imaging.High-resolution images usually contain more image information,richer details,and pleasing visual effects.In many application scenarios,we need to obtain high-resolution images to meet task requirements.For example,in medical imaging,satellite imaging,remote sensing,computer vision and other fields,image super-resolution has a wide range of important applications.However,due to the limitations of the sampling equipment,the loss of the transmission process,environmental interference and other factors,we cannot obtain ideal high-resolution images.Low-resolution images lack a lot of pixel information,and each low-resolution image usually corresponds to many high-resolution versions,which is a typical ill-posed problem.Therefore,the task of image super-resolution is very challenging and difficult,and it is also a research hotspot of many scholars.Due to the important application value of image super-resolution,the demand for high-resolution images is becoming more and more urgent,which has accelerated the development of image super-resolution technology.Because high-precision sampling equipment and processing equipment are expensive,their application range is limited.Compared with the hardware method,the software has superior reusability and portability,which can effectively expand the scope of application and reduce the cost.In recent years,a large number of image super-resolution algorithms have been proposed,and the results of image super-resolution have been greatly improved.However,many single image super-resolution algorithms usually have their inherent limitations,and the results obtained still need to be further improved.The interpolation-based method usually lacks high-frequency information due to the limited fitting accuracy of the interpolation function;the reconstruction-based method relies excessively on prior information,and the effect is poor when the prior information does not match the image features;The learning-based method usually relies on external examples and have high complexity.To this end,this paper makes use of the non-local self-similarity,cross-resolution similarity,and global structure sparsity of images,and proposes a new single-image super-resolution method without relying on external examples,which aims to improve the non-local self-similarity,cross-resolution similarity,and global structure sparsity.Learn the super-resolution accuracy of the image in the case of an example.First,we obtain the initial high-resolution image through feature-constrained polynomial interpolation.Then,we use the lightweight database constructed by the input image itself to perform cross-resolution learning to predict the high-frequency information missing in the interpolation result.Finally,the residual filtering proposed in this paper removes the noise introduced in the process of interpolation and cross-resolution learning.The feature-constrained polynomial interpolation we proposed can be combined with other image super-resolution algorithms.Cross-resolution learning can be used to predict the missing high-frequency information of super-resolution images obtained by other interpolation methods,and residual filtering can be used as the post-processing of other algorithms.In the experiment,we compare our method with many of the most classic image super-resolution algorithms.The experiment content includes:1.Numerical accuracy;2.Visual effect;3.Validity verification;4.Scalability verification.Extensive experiments have verified the superiority of the HR images obtained by our algorithm in numerical accuracy and visual effects,and has validity and scalability.