Cascading-learning-based Single Image Super Resolution

High-resolution images take advantage of more pixels per unit area,enriching the texture of the image and revealing more image information.However,in actual situations,satisfactory high-resolution images cannot be obtained due to the limitations of the imaging devices themselves and various external factors.Image super-resolution reconstruction techniques can be used to obtain higher resolution images by processing existing low-resolution images.At present,the learning-based single image super-resolution reconstruction methods are the mainstream methods in the field of image super-resolution reconstruction.Several common learning-based methods are studied and analyzed in this paper.The main research contents are summarized as follows:The neighborhood embedding method may have the following problems when the training set images are selected: if the training set contains few images,the input image may be inconsistent with the training set images,which may affect the reconstructed image quality.If there are more images in the training set,the search workload of similar image patches will increase,and the efficiency of the algorithm will be reduced.In order to solve this problem,a single image super-resolution algorithm based on self-similarity and non-local constraints is proposed.The algorithm does not rely on the training set images,instead,it takes the self-similarity of the image as the prior knowledge to directly adopt the strategy of small-scale magnification and multiple reconstructions on the input image.The reconstruction process adopts the neighborhood embedding method.When searching for similar image patches,the search range is adaptively selected according to the statistical characteristics of a single image to improve the search efficiency.Finally,the regularization term is constructed using the non-local characteristics of the image to synthesize the final high-resolution image.Experimental analysis and comparison show that the proposed algorithm has better reconstruction results.Regression-based methods usually require large number of training set images,and through learning many local mappings on the training set to enrich the texture details of the reconstructed images.There are the following problems: too many training set images will lead to training set redundancy and high computational resource requirements.Whenlearning the local mappings,the problem that the actual local high-resolution feature distributions are inconsistent with the corresponding local low-resolution feature distributions is often neglected.Moreover,most of the algorithms only learn simple local mappings,and do not consider the overall characteristics of the training set,and cannot accurately describe the relationships between high-resolution images and low-resolution images.In order to the above problems,a single image super-resolution algorithm based on pre-classified cascade mappings is proposed.This algorithm pre-classifies the training set images to reduce the redundancy.And local high-resolution feature distributions are used to correct the local mappings.Furthermore,the learned local mappings are applied to the lowresolution images of the training set to obtain the reconstructed images,and the overall relationship between all reconstructed images and the high-resolution images is established in the training set,and the overall mapping coefficient is obtained by using the multivariate linear regression model.Finally,the high-resolution reconstructed image is obtained by cascading the local mappings and the overall mapping coefficient.The experimental results show that compared with some state-of-art algorithms,the proposed algorithm can recover more texture details and has higher reconstruction quality.