Hyperspectral Image Compression Based On Sparse Decomposition

Hyperspectral images have great application significance for they can be used for the pixel-level or even sub-pixel identification effectively. However, its high resolution ratio takes larger amount of data as a cost, which brings a big problems for transmission and storage. So the research of high performance compression algorithms has important application significance.This paper proposed a lossy compression scheme which used sparsify and sparse decomposition together. What's more, it also builded another system as compared, which used SPIHT algorithm based on DWT after sparsify.Firstly, the image's sparsity means it has far more contour information than detailed. The model used in this paper can effectively sparse the hyperspectral images. The test data shows that it makes sparse rate (zero pixel ratio) increased from less than 0.6% to 55% above. In addition, the sparse process can also remove the spectral correlation more than 94%.Secondly, the paper used two ways to compress the spared images. The first one used signal sparse decomposition algorithm to decompose them, then used fixed-length quantization method encoding the atoms selected by MP algorithm. Another one used SPIHT algorithm based on lifted 5/3 wavelet. The results show that the performance of sparse decomposition is worse than SPIHT in near lossless compression (PSNR> 40dB). But in the high-rate compression, the advantage of sparse decomposition is obvious. The average PSNR is 31.12dB at the compression ratio of 102.40. This performance is better than many references'.Finally, the paper analyzed the differences of the two algorithms, and demonstarted the ROI scheme which used both of them. In this scheme, the interest region (40% of the whole image) used SPIHT algorithm with PSNR over 90dB, and the background region used sparse decomposition algorithm with PSNR over 30dB. The result is that the compression ratio can reach about 4.4.