Research On Super-resolution Algorithm Based On Sparse Representation And Feature Fusion

In recent years,with the rapid development of communication technologies and the advent of the information age,high-resolution images have become increasingly prominent in various fields such as video surveillance,medical imaging,and remote sensing satellite imaging.The image super-resolution reconstruction is based on one or more low resolution images to reconstruct the high resolution image.Which breaks through the inherent limitations of existing image imaging devices and achieves better application of high resolution images.Therefore,super-resolution reconstruction has important research significance.The traditional learning based super-resolution reconstruction algorithm is complex,the computational complexity is large and the prior knowledge is single.In view of this problem,we have studied the two aspects of improving super-resolution reconstruction speed and the fusion of internal and external features.The main achievements of this paper are as follows:(1)For the fast single image super-resolution via self-example learning and sparse representation is used to train dictionary by using input images.The reconstructed image is smooth and artifact,and the reconstruction speed is slow.A fast image super-resolution for residual dictionary learning algorithm is proposed.Based on the concept of residual image,dictionary training is done by high-frequency residual images extracted from external high-resolution images.The dictionary training time and the amount of calculations are reduced,and the performance of the high frequency residual dictionary is better.In order to improve the speed of reconstruction,the Cholesky decomposition method is used to simplify the calculation of inverse matrix in the orthogonal matching tracking algorithm,and the sparse coefficient can be quickly solved while reducing the amount of calculation.The high resolution image is reconstructed by the obtained sparse coefficient and the high-frequency residual dictionary,and the iterative back-projection method is used to further improve and reconstruct the reconstructed image.Experiments show that the algorithm has fast reconstruction speed and good image effect.(2)For the shortcomings of image super-resolution reconstruction with single prior knowledge constraints,this paper proposes image super resolution based on fusing internal and external features.The pseudo high-resolution image is obtained by bicubic interpolation of the input image,and the high-frequency detail of the pseudo high-resolution image is enhanced by the deep neural network(ARCNN)and artifacts are removed to obtain a high-frequency residual map.At the same time,the pseudo high-resolution image is input into a deep neural network(SCN)with sparse priors.The residual dictionary trained by the external image set quickly reconstructs the initial high-resolution image.The fusion model based on convolution sparse coding is used to fuse the high frequency residual image and the initial high resolution image,and the super-resolution reconstruction of the image is effectively realized.In order to improve the visual effect of the reconstructed image,an iterative back-projection method is used to further improve the image quality.Experiments show that the algorithm improves greatly in peak signal-to-noise ratio and image visual quality.