Research On Learning-based Image Super-resolution Reconstruction Algorithms

With the development of information technology,people pay more and more attention to the vividness and visualization of information.High-quality images just can meet this demand.But in the existing imaging system,the resolution of the image can not meet the actual demand because of the limitation of hardware and the uncertainty of the outside world.Seeking methods to obtain high quality images from the hardware requires a high cost.Therefore,the method to solve this problem from software came into being.Compared with the traditional super-resolution reconstruction algorithm,the learningbased image super-resolution reconstruction algorithm is a research hotspot recently.On the basis of studying the current learning algorithm,the existing problems in the existing algorithms are improved.Then,based on the shortcomings of the existing algorithms,image super resolution reconstruction based on the extreme learning machine is proposed.The main research contents and innovations are as follows:(1)The dictionary construction is the key of super resolution image reconstruction method based on learning.Aiming at the problem that the existing dictionary learning algorithm is insufficient,and it can not obtain enough a priori information to reconstruct the low resolution image,so we propose an improved dictionary learning algorithm.The improved dictionary learning algorithm is used to update the dictionary matrix and nonzero elements in the sparse coefficient matrix.In combination with the idea of the joint dictionary,the improved algorithm has a smaller sparse error and the dictionary atoms are better and the dictionary converges faster.(2)Aiming at the problem that sample image blocks contain more information and the dictionary training time is too long,a super-resolution method of low rank matrix combined with improved dictionary is proposed.The image is decomposed into low rank part and sparse part,which can make use of the feature information more effectively and obtain high quality super resolution image.(3)In order to solve the problem of the blocky effect in sparse representation of image reconstruction and the lack of detailed information of the reconstructed image,the super-resolution reconstruction of the improved wavelet domain is proposed.The more compact dictionary is trained on the single feature space,which can better restore the image of the local texture and edge.Experiments on low-resolution images with noise have yielded good results.(4)Image sparse representation and dictionary learning algorithm,learn from the idea of machine learning algorithm,we use the extreme learning machine with higher efficiency and less parameters to achieve image super-resolution reconstruction.On the basis of dictionary learning,the theory of extreme learning machine is applied to image super-resolution reconstruction,and the extreme learning algorithm is used to optimize the image super-resolution reconstruction.Finally,the experimental results show that the proposed super-resolution reconstruction models have a better performance in terms of objective evaluation and subjective vision.