Research On The Application Of Super-Resolution Method Based On Dictionary Learning

Due to the limitation and expensive cost of imaging equipment,super-resolution methods are widely used to reconstruct high-resolution images from low-resolution images.Among various super-resolution methods,interpolation-based super-resolution approaches are simple,fast and suitable for real-time applications,but the resultant image is blurry and with jagged edges.Reconstruction-based super-resolution algorithms produce sharp edges and reduce image noises.However,when the value of upscale factor is large,these kind of methods cannot effectively restore the image details which were lost during the degradation process.Hence,the quality of the reconstructed images is unsatisfactory.By comparison,learning-based superresolution methods utilize a large amount of high and low resolution samples as prior information,so the reconstructed high-resolution images contains more detail information and the reconstruction quality is better.Therefore,our thesis studies the applications of superresolution technology based on dictionary learning.In our paper,we first introduce several classical algorithms in several types of superresolution methods and explain their basic theories,implementation details,strengths and weakness.Then,we introduce the theory and applications of transfer learning.Hence,to tackle the weaknesses of generic super-resolution methods,we propose two application-based superresolution methods.One is based on transfer learning theory for content specific applications and another one is based on spatial-adaptive learning for barrel lens distorted images.For our super-resolution method based on transfer learning,the first step is transferring supplemental images from other standard datasets and combining them with the original training data to construct a new dictionary.The supplemental images provide more training samples to support transfer learning at the stages of sample selections.Then,we select nearest neighbours from training samples for each dictionary atom.According to these neighbours,projection matrixes are computed to reconstruct high-resolution images.In the experiment validation part,four experiments are conducted on three datasets to explore which types of images can provide more suitable information for super-resolution tasks.The results indicate that our transfer learning approach improves the PSNR and SSIM values,and the quality of the reconstructed images is obviously higher when the transferred images contain similar content to the original ones.For our second super-resolution method for barrel lens distorted images,the input lowresolution barrel distorted images are firstly un-distorted to restore the original rectilinear shape.Then,the popular A+ method is applied to conduct the first reconstruction procedure to form the reconstructed high-resolution image.After that,for the high-resolution image,we use postprocessing procedure to classify the image patches,and compute coefficient matrix for each class.These matrixes are then used to refine the image patches for reconstructing the final highresolution results.Experiment results prove that the proposed approach can improve the images in terms of PSNR and SSIM values and achieve better visual quality for barrel lens distorted images.