Learning-based Single Image Super-resolution Reconstruction Method And System

Many images obtained daily have low resolution.Low-resolution images with a lot of details lost affect information understanding and analysis,which causes image processing challenging.Super-resolution image reconstruction technology can improve the image resolution using software without changing various imaging factors,but the super-resolution image reconstruction is an ill-conditioned inverse problem,which requires the use of appropriate prior knowledge to constrain the solution.Super-resolution methods based on learning automatically learn features from training samples and thus obtain better reconstruction results than other methods.This study focuses on the study of single-image super-resolution reconstruction based on learning.The main innovative work is as follows:(1)Considering many current super-resolution methods use single global dictionaries,which are difficult in reconstructing image blocks with different structures,a multi-dictionary image super-resolution method based on a student t-distribution hybrid model is proposed.Specifically,the stationary wavelet transform is used to extract image features firstly.The student-t distribution mixture model(SMM)is adopted to group image patches into clusters with each corresponding to a different pair of dictionary learned by the K-SVD method secondly.With multiple types of dictionaries,the high resolution image is reconstructed using weights that are formed by the probabilities based on likelihood between the image and different clusters,and then globally optimized using an improved iterative back projection(IBP)method finally.(2)For the problem that current image super-resolution algorithms based on convolutional neural networks cannot take full advantage of high-frequency information in features and cannot achieve non-integer scale of super-resolution reconstruction,an arbitrary scale image super-resolution method combining attention mechanism is proposed.In this study,Global residuals are utilized to reduce the difficulty of convolutional neural network learning,and the local residuals are combined with attention mechanisms to improve the training efficiency of the network.Finally,the meta-learning up-sampling structure is applied as the amplification module of the network,which improves the quality of image reconstruction and allows for arbitrary scale reconstruction of the image.(3)An image super-resolution reconstruction and evaluation system is designed and implemented.The system integrates multiple commonly used image super-resolution reconstruction algorithms such as bicubic interpolation,as well as the multi-dictionary based and deep learning methods proposed in this paper.The main functions of the system include the selection of different super-resolution reconstruction methods to reconstruct,and the qualitative and quantitative evaluation of the reconstructed results.