Research On Key Technologies Of Machine Learning-Based Image Super-Resolution

The goal of image super-resolution(ISR)is to reconstruct high-resolution image from low-resolution image.After more than 40 years of development,ISR technology has gained great progress and has been widely used in the fields such as medical imaging,remote sensing and security monitoring.Among the existing ISR methods,interpolation-based methods usually estimate unknown pixel values on high-resolution image raster through known pixel values on low-resolution image raster.Commonly used interpolation methods include polynomial interpolation methods and edge-directed interpolation methods.Interpolation-based methods have low computational complexity,but they can easily cause edge smoothness,blurring and aliasing in the reconstructed high-resolution image.Reconstruction-based methods can break through the defect of traditional interpolation-based methods by using some prior knowledge which is related to image degradation process to establish ISR model.However,the application of reconstruction-based methods is limited by the difficulties in manually descripting the features of image priors.Learning-based methods learn the mapping relation between high-and low-resolution images from training samples by applying machine-learning.The learned mapping relation is then used in reconstructing high-resolution image.In practice,the application scenarios of ISR vary in performance requirements such as computational complexity,image quality and hardware cost.In this dissertation,author focuses on both researching sparse representation-based fast ISR methods for low time consumption and low hardware cost scenario and researching deep learning-based methods for high-performance computing scenario.The main contents and contributions are summarized as follows:(1)A hierarchical regression model based on clustering and sparse representation is proposed to learn the mapping relation between high-and low-resolution images from training data.Firstly,the sample image patches are classified into different clusters.Then a pair of high-and low-resolution dictionaries is learned for each cluster by applying dictionary learning algorithm.Finally,a series of mapping matrixes related to every pair of dictionary atoms are learned by applying linear regression.Traditional fast super-resolution methods,such as multiple linear mapping model and anchored neighborhood regressioin model,only exploit single feature of clustering or sparse dictionary atoms while constructing mapping functions,which makes the functions too simple and results in difficulties in restoring high frequency details.The proposed method have solved these problems.Experimental results show that the accuracy of mapping relation modeling is improved by hierarchical regression,which leads to improved ISR quality.(2)An L2,1-norm regularized collaborative sparse dictionary learning model is proposed,in which the structure consistency of clustering image patches is used as prior information to constrain the selection of dictionary atoms.The dictionary atoms are used to replace the original image patches as regression sample to construct mapping model for more accurate and robust ISR.A two-step iterative algorithm is proposed to solve the collaborative sparse dictionary learning problem.The iterative algorithm uses randomly selected original image patches as initial dictionary,and then the dictionary and sparse representation matrix are updated by alternating iterations until the convergence condition is satisfied.Experimental results show that the proposed method improves ISR quality by improving the selection of sparse dictionary.(3)A deep learning-based method named multi-channel densely connected convolutional network is proposed for ISR.The proposed method combines empirical mode decomposition(EMD)with deep learning.EMD is used to separate different frequency components from original input image.Then multi-channel convolutional networks with dense connections are used to learn deep features from these separated frequency components.Finally,all these features are fused to obtain desiered high-resolution image.When compared to single channel network,the proposed multi-channle network can improve the expression of different frequency characteristics existing in super-wide spectrum images if network depth becomes deeper.Experimental results show that using EMD to separate different frequency components of input image simplifies the complexity of image expression and improves the accuracy and efficiency of the neural network in image feature learning.(4)A dual path hybrid network is proposed for ISR.The proposed method improves deep residual network by adding a densely connected path to it.Deep residual network can reuse feature information at a relatively low redundancy,and densely connected network can learn new feature information from resused information.The hybrid structure can help fully utilize the adavantages of both deep residual network and densely connected network to improve the efficiency and accuracy of image feature learning,which will finally improve the ISR quality.