The Research Of Spectral Feature Based Hyperspectral Image Compression

HyperSpectral Image (HSI) can get the spectral information and the spatial information of the targets simultaneously. Because of the special advantage, hyperspectral image obtains more and more applications in wide areas. But the spectral information of hyperspectral image brings a drawback of huge data, which is inconvenient to storage and transmission. Hyperspectral image is a kind of 3D images, which is different from common 2D image and video image. The compression of hyperspectral image is a new problem in image compression area. So how to find proper compression methods based on the characteristics of hyperspectral image is important.First, to make the research successful, the characteristics of hyperspectral image are analysed. Hyperspectral image has spectral correlation, which is different from common digital image. The spectral correlation is stronger than spatial correlation. Through the analysis of hyperspectral image dimensionality, we concluded that the hyperspectral image spectral spaces are mostly empty. It's the theoretic basis of HSI compression. HSI has hundreds of bands, which have different importance. It can be seen from the entropy of different bands. From the analysis, we can see that the particularity of HSI is the spectral dimensionality. The following compression methods will focus on the point. Thus the HSI compression research around the spectral feature will get better results.Second, the spectral characteristics based fast Vector Quantization (VQ) methods are proposed for HSI compression. Vector quantization is an effective lossy compression method, especially in high compression ratio. It's the mainstream method of HSI compression. But the encoding complexity of vector quantization is high. When vector quantization is applied to HIS, the practicality greatly reduced. So the emergency problem of VQ based HSI compression is the fast algorithm. To solve the high computing complexity of VQ, the spectral feature selection based VQ and the spectral feature transformation based VQ are proposed. The fast algorithms can reduce the computing time, while the reconstruction SNR is slightly reduced. So it can solve the problem of high computing complexity.Third, optimal prediction based on linear model for HSI compression is proposed. The vector quantization does not have superiority in the non-great compression ratio, but the prediction technology has its merit. Traditional prediction based method is the mainstream method for HSI compression too, but it has a drawback of low prediction efficiency. To solve the problem, a linear model for HIS is established. The model can realize the optimal prediction under the meaning of SNR, so it can get better reconstruction result. Two hyperspectral images are tested at the same CR. The reconstruction signal to noise ratios are improved 1.5dB and 5.5dB using optimal prediction.Last, the concept of interested information in HSI compression is proposed, and the corresponding method is proposed too. The information of interest in HSI contains two parts: the Region Of Interest (ROI) and the Band Of Interested (BOI). The small target preservation is very important to HSI application. To preserve the spatial information of the small target, a SPIHT based algorithm for ROI based compression is proposed. The band information preservation is a particular problem of HSI compression, and the band of interest is connected to specific application. A band of interest preservation algorithm is proposed for classification. The algorithm can get better classification accuracy while the SNR is slightly reduced.In this dissertation, several algorithms for HSI compression are proposed. These algorithms focus on the spectral characteristeic, which is the unique feature of HSI. Some analysis and experiments have been done to evaluate the algorithms.