Research Of Algorithms Of Image Super-resolution Reconstruction

With the further development of image processing technology,the requirement of image resolution in practical application is increasing,however,the quality of images will be affected by some hardware devices,environments and other random factors in the process of image acquisition,transmission,reproduction and display,etc.Therefore,the image super-recolution algorithm has become a hot issue in current research,which uses software algorithms to improve image resolution without increasing the cost of hardware imaging equipment.The image super-resolution reconstruction algorithm uses high-frequency complementary information learned from low-resolution image sequences to reconstruct the corresponding high-resolution image through digital image processing algorithm,which has some advantages in cost and fliexibility.Hence,the image super-resolution reconstruction algorithm plays an important role in image processing applications,such as satellite remote sensing,medical imaging,surveillance,astronomical observation and biological information identification,etc.This dissertation focuses on the enhencing quality of the reconstructed image in the super-resolution algorithm.This paper proposed an image super-resolution reconstruction algorithm based on K-SVD dictionary learning,an image super-resolution reconstruction algorithm based on adaptive regularization selection,an image super-resolution reconstruction algorithm based on Fields of Experts prior model and an image super-resolution reconstruction algorithm based on convolution sparse auto-encoder.The main contributions are as follows:(1)Aiming at the problem of long training time in semi-coupled dictionary pair algorithm,an image super-resolution reconstruction algorithm based on K-SVD dictionary learning is proposed.The proposed algorithm first uses the orthogonal matching pursuit algorithm for sparse coding,and then uses K-SVD algorithm to update the dictionary atom and its corresponding sparse representation coefficient at the same time.Finally,the coefficient mapping function is also trained.The proposed algorithm decreases the number of iterative in the training phase,thereby reducing dictionary training time.The simulation results show that the proposed algorithm reduces the simulation execution time by three and four times on average in different dictionary sizes and improves the quality of image reconstruction.(2)Aiming at the lack of adaptiveness of regularization parameters in local autoregressive and non-local similarity sparse representation algorithm,this paper proposes an image super-resolution reconstruction algorithm based on adaptive regularization parameter selection.The proposed algorithm first uses the image local variance to set the threshold,which divides the image into smooth regions and non-smooth regions,and then adaptively selects regularization parameters in different regions of the image.So that the parameters are adaptive to the structural features of different regions.The simulation results show that the proposed algorithm achieves better reconstruction results than the empirical selection of regularization parameters.(3)Aiming at the noise problem of non-local centralized sparse representation algorithm,an image super-resolution reconstruction algorithm based on Fields of Experts is proposed.The proposed algrithm first uses the Fields of Experts model to learn the prior knowledge of the whole image and learns all the parameters from the image training sets.The super-resolution algorithm updates the prior constraint and super-resolution reconstruction simultaneously.The simulation results show that the proposed algorithm enhances the effect of image super-resolution reconstruction and de-noising.(4)Aiming at the problem of feature maps accuracy in convolutional sparse coding algorithm,an image super-resolution reconstruction algorithm based on convolutional sparse auto-encoder is proposed.The proposed algorithm first uses the sparse auto-encoder to pre-train the feature maps of input images,and then uses the obtained low-resolution and high-resolution feature maps to learn the corresponding filters and mapping function,finally the high-resolution image can be estimated using the high-resolution filters and feature maps,which not only utilized the consistency constraint but also improve the sparse coding ability.The simulation results show that the propose algorithm effectively enhances the quality of reconstructed images.