Block-sparse Representation Based Compression For Hyperapectral Images

With the progress of remote sensing imaging technology,the amount of hyperspectral image data has increased dramatically.This is not helpful for image storage and transmission,and also can cause waste of resources.Therefore,the compression of hyperspectral images is of great importance.The traditional hyperspectral image compression methods are mainly based on prediction and transformation,following the Shannon Nyquist sampling theorem,with the problem of large sampling rate,and the structure feature of the image is not considered in the compression and reconstruction.In this paper,compressed sensing is utilized to compress and sample simultaneously and block-sparse dictionary is adopted to build the structural sparse representation model.Structural compression methods based on the block-sparse dictionary are proposed for the two-dimensional and three-dimensional compression models of hyperspectral images.Firstly,a hyperspectral image two-dimensional compression and reconstruction method based on block-sparse dictionary is proposed.In the encoder,the measurements is obtained by the observation matrix,and the image is reconstructed by the structural adaptive block-sparse dictionary in the decoder.The reconstruction time of this method takes a long time,but the coding efficiency has been improved by the block-sparse coefficient coding algorithm.The experimental results show that the selection of block-sparsity value is related to the number of ground features of hyperspectral image and the size of the atom block,so an appropriate block-sparsity should be selected according to the actual situation.Compared with non-learning dictionaries,unstructured dictionaries and fixed structure dictionaries,this method improves the compression performance.Then,in order to solve the problem of the above-stated method that the computational complexity of the decoder is high and the compression rate is not very low,a two-dimensional compression and reconstruction method based on block-sparse coefficient coding is proposed.The original image data is sparse coded by the block-sparse dictionary in encoder,and then three matrices are obtained by the block-sparse coefficient coding algorithm.The data is reconstructed in the decoder.The advantage is that it shortens the reconstruction time and reduces the compression ratio.It can achieve high reconstruction quality without disrupting the structural characteristics of the image.This method is applied in the case of having abundant resource in the encoder,which can reduce the compression ration and guarantee the accuracy of reconstruction.Lastly,the hyperspectral image is regarded as a three-dimensional tensor signal to realize the spatial-spectral joint compression in order to solve the problem that the separate operation in the two-dimensional compression model destory structure feature of the image,and a hyperspectral image compression method based on the multidimensional block-sparse representation and dictionary learning is proposed in this paper.Considering the respective feature structure of the hyperspectral tensor on each mode,dictionary learning algorithm is adopted to train three dictionaries on each mode.These dictionaries are applied to build the block-sparse model of hyperspectral image.Then,based on the Tucker decomposition,the spatial and spectral information is compressed simultaneously to get the measurement and the measurement is also a tensor.Finally,the structural tensor reconstruction algorithm is utilized to recover the hyperspectral image.The biggest advantage is that the block-sparse tensor reconstruction algorithm avoids complex nonlinear reconstruction,improving the computation efficiency.And it provides a flexible mechanism to adjust the sampling rate of spatial and spectral domain according to the practical situation.