Research On Hyperspectral Image Coding

Hyperspectral imaging, known of its nanometer scale resolution, is capable of providing rich spatial and spectral information of target scene. For this reason, it has been widely used in both military and civilian fields. The challenge that this technique faces is high performance compression of massively large sized data sets. Therefore low complexity, low memory requirement and high performance hyperspectral image coding method become much meaningful. In this dissertation hyperspectral image coding method is studied, and the major work of this dissertation consists the following:First, the spatial and spectral correlation of hyperspectral image is investigated. Two-dimensional spatial decorrelation method which is based on wavelet transform and one-dimensional spectral decorrelation method which is based on Karhunen-Loève transform (KLT) are proposed. The symmetric factorization of biorthogonal wavelet based reversible integer spatial combinative lifting algorithm is studied. Pseudorandom uniform sampling is used to reduce the calculation of the covariance matrix in KLT, and the matrix factorization of transform matrix is used to research on reversible integer KLT. The experimental results demonstrate that the three-dimensional decorrelation method can effectively decorrelate the spatial and spectral correlation of hyperspectral image, and has low complexity as well.Secondly, based on the analysis of the transformed coefficients, the three-dimensional significance tree is constructed from three-dimensional spatial-spectral zero tree, and the node of significance tree provides significance information of the coefficients. The three-dimensional hyperspectral image coding using significance tree, both with list and without list are proposed. The node of significance tree, three-dimensional context of coefficients and reversible integer decorrelation transform are combined together to achieve lossy to lossless progressive coding method of hyperspectral image. The experimental results demonstrate that the coding using significance tree without list has good coding performance and can reduce the complexity and memory requirement of encoding and decoding process as well.The theory of arithmetic coding and the renormalization are then studied. In order to solve the problem during the renormalization, an improved fast arithmetic coding is proposed. It does renormalization in bytes style, and reduces the comparing operation of the width of coding range in the encoding and decoding process, and outputs codes in bytes style. In addition, it deals with the carry propagation problem more effectively. The experimental results show that the proposed method has faster speed of encoding and decoding process. Then the proposed method is applied into hyperspectral image coding to simplify the bitplane encoding and decoding process.Finally, the multiple bitplanes interleaving algorithm in region of interest (ROI) coding is analyzed, and the calculation of three-dimensional mask of ROI and the nodes of significance tree is studied. The three-dimensional ROI coding algorithm based on the multiple bitplanes interleaving algorithm and the embedded bitplane coding algorithm using significance tree is proposed. At low and medium bit rate, this method has the capability of effectively protecting the quality of ROI by controlling the relative significance and embedded form of ROI and region of background.In summary, key techniques of hyperspectral image coding are studied, including decorrelation transform, bitplane coding, arithmetic coding and ROI coding. A lossy to lossless progressive hyperspectral image coding method with low complexity, low memory requirement and high performance is proposed.