Remote Sensing Image Super Resolution Reconstruction Based On Sparse Dictionary Learning

The spatial resolution of image is a key technical index in the design of remote sensing imaging system.Higher resolution means higher data volume and helps us to get more useful information.At present,remote sensing technology can provide various ground observation data quickly and effectively,and become the main information source in topographic mapping,urban construction,road construction,military reconnaissance and other fields.However,due to the large distance between the satellite remote sensing imaging system and the target,the target can not be accurately identified,which makes the object recognition degree after imaging low and the details less.Therefore,in view of the characteristics of low-resolution remote sensing images,such as similar image blocks in non-local areas,rich texture and obvious edge structure,different methods are adopted to achieve single-frame super-resolution reconstruction of remote sensing images.Combining sparse representation theory and dictionary learning method,this paper focuses on single remote sensing image super-resolution reconstruction technology,including:(1)Analyzing the various influencing factors in the imaging process of satellite remote sensing,giving the degradation process of remote sensing image,Then,the sparse representation theory is elaborated.Finally,three typical super-resolution reconstruction methods,commonly used dictionary learning algorithm and image quality evaluation index are introduced.(2)An image super-resolution algorithm based on non-local self-similarity dictionary learning is proposed.For satellite remote sensing can not obtain high-definition images due to external factors,the current commonly used sparse representation based algorithm for reconstruction of details is not good.Considering the similarity of two images in different regions of remote sensing image,this algorithm introduces the non-local self-similarity constraint to the sparse representation model of a single remote sensing image.Firstly,according to K-Means algorithm,k points are randomly selected from high-resolution image blocks as the initial centroid,fully considering the similarity of each image block in the process of image reconstruction,and according to the non-local self-similarity constraint.Self-similarity clustering into k clusters,and then using singular value(SVD)decomposition to achieve PCA dictionary.The experimental results show that the algorithm achieves good results in remote sensing image reconstruction,and the details are clearer.(3)An image super-resolution algorithm based on multi-dictionary learning is proposed.Some remote sensing images have complex structures,different features,poor expressive ability of singletons,difficult to accurately represent different types of image blocks,and long computing time.The algorithm uses MCA(Morphological Component Analysis)decomposition method to get the structure layer image and texture layer image;then uses different decomposition methods for different image layer features to increase the accuracy of feature matching;in the dictionary learning stage,trains different image layer height and low resolution dictionary pairs;and then calculates the initial image and the original high score after different image reconstruction.The residual images between the resolution images are trained to further reconstruct the high resolution images.Finally,the structural images are restored by using total variation to obtain the final high resolution remote sensing images.The experimental results show that the structure and texture of the reconstructed image are clearer,and the quality of the reconstructed image is better,both in vision and in evaluation index.