Sparse Representation And Its Application In Cloud Image Super-resolution

Satellite cloud image is a kind of image data which is received by the imaging sensor on the satellite base on the observation of the earth surface,it is the core tool of weather analysis,meteorological research and disaster monitoring.Due to the constraints of the launch technology,satellite weight,sensor accuracy and other factors,the cloud image resolution often does not meet the actual engineering needs.The using of more accurate satellite imaging sensor can improve the resolution of the cloud image,but it is limited by cost and technology,in order to solve this problem,we use computer software to enhance the resolution of cloud image.In view of the flourishing development of sparse representation theory in signal processing,In this paper,the sparse representation and its application in cloud image super-resolution are studied,the main contents are as follows:(1)A cloud image interpolation reconstruction algorithm is proposed based on morphological component analysis.The cloud image is composed of several morphological structures,each of which has different image features.In this paper,the low resolution cloud image is decomposed into smooth component and texture component by morphological component analysis(MCA).Using the good characteristics of stationary wavelet transform(SWT)in structural feature expression,and the advantages of Contourlet transform in image texture feature representation,the smooth component and the texture component are reconstructed by using the stationary wavelet transform and Contourlet transform respectively.In the process of reconstruction,in the light of the problem of low frequency energy loss by wavelet interpolation,the effect of stationary wavelet interpolation is improved by using the method of first interpolation and then transform.Finally,the reconstructed smooth component and texture component are added to obtain the high resolution image.The results show that the proposed method can solve the problem of smoothing and artifact in the reconstructed cloud images,and it is superior to other interpolation algorithms in visual effects and PSNR.(2)A cloud image super-resolution reconstruction method based on topic learning and sparse representation is proposed.Firstly,the cloud image is decomposed into smooth and texture parts by the bilateral filter,and the texture part is regarded as a training sample composed of several “documents”.Latent semantic features of "document" are extracted by probabilistic latent semantic analysis(pLSA),to discover the inherent “topics” of “documents”.On this basis,the improved KSVD method is used to train the high and low resolution dictionary for each topic.In the reconstruction phase,the texture component is used to achieve the super resolution reconstruction of each topic block in the corresponding dictionary by the sparse representation method.The smoothing component is reconstructed by the direct interpolation method.Finally,the reconstructed smooth and texture components are added to obtain high resolution cloud images.The experimental results show that the reconstructed cloud image is better than the traditional method in visual effect and PSNR.