Complex Image Compression And Classification Via Reversible Integer Karhunen-loeve Transform And Feature Caling Based KFDA

Digital images have been playing an increasingly important role in information areas. The digital image storage and transmission is facing big challenges. Image compression technique can help to reduce the burden of storage and transmission of the huge image data volumes. Especially, progressive lossy-to-lossless compression obtains much attention since it processes a great amount of application values. For hyperspectral images, lossless compression is important to preserve the value of original information. Progressive compression technique based on shape-adaptive integer transforms and automatic ROI extraction system can support lossless compression of ROI areas and lossy compression of BG area, the technique can retain the important information about the ROI and reduce storage space. The theory and techniques for image compression are researched from three aspects, disciplines of reversible integer transform, image compression using reversible integer transform and image compression based on region of interest extraction in this dissertation.The main content of this dissertation is summaried in five points as follows:1. We makes detailed analysis about some transform methods which are always used in image compression, based on the basic discipline of floating-point transforms. And then we discussed about lifting scheme and matrix factorization theory which forms the basis of the reversible integer transform study. We propose a progressive compression method for3D color image based on integer reversible implementation of KLT (Integer Karhunen-Loeve Transform, IntKLT). Compared with conventional KLT-based algorithms, the improvements consist of two aspects. Firstly, we decompose KLT matrix into triangular elementary reversible matrices (TERM) using quasi-complete pivoting matrix factorization method, and reversible integer transform is realized using this method. Experiments testify that the efficiency of this integer transform approximates that of original floating-point transform very well. Secondly, we encode transformed coefficients using set partition embedded block algorithm which proves to be an efficient codec. Compared with conventional KLT-based compression algorithms, the proposed method can realize both lossy compression and lossless compression. Besides, reversible integer transform is realized using quasi-complete pivoting matrix factorization method and low computational complexity.2. We study about3D color image compression methods based on low-complexity reversible integer KLT and dynamic programming. The computational complexity of IntKLT is very high and the approximation ratio of KLT is unsatisfied. In order to improve the approximation ratio and the computational complexity, a new progressive compression method for3D color image based on low-computation integer reversible implementation of KLT and dynamic programming is designed. Compared with IntKLT-based compression algorithms, the proposed method can realize both low computational complexity and high accuracy. Besides, the approximation ratio of KLT is realized using dynamic programming method.3.In hyperspectral image compression based on spectral KLT, since the huge amount of hyperspectral image data cannot all be processed at once, data need to be blocked, using the KLT in the spectral domain. The traditional block-based KLT leads to obvious distortion of spectrum curves in the block boundary. We propose a new transform scheme called the lapped Karhunen-Loeve transform (KLT) for hyperspectral image compression. The proposed lapped KLT transform is composed of a traditional KLT and pre-filter. The pre-filter considers the correlation of the adjacent spectral block, and reduces the correlation between adjacent blocks. The lapped KLT thereby solves the problem of block boundary distortion caused by KLT. The experimental results show that for the spatial domain of an image using the same transform and the same coder with a medium to low rate, our lapped KLT method outperforms KLT alone in the spectral domain.4. We focus on ROI segmentation and classification. We propose a classification method for hyperspectral images based on feature scaling for kernel fisher discriminant analysis (FS-KFDA) and feature extraction. Shape-adaptive transform and coding is the important components of ROI compression, people usually pay most of their attention to only part of shape-adaptive transform and coding. But ROI extraction play an important role in ROI compression, it can affect performance of ROI compression, therefore, ROI extraction should be considered more severe than elsewhere. In this paper, we proposed a new ROI extraction algorithm based on FS-KFDA and feature extraction. Starting from spectral imaging mechanism, we make full use of spatial and frequency feature. Besides, FS-KFDA is very well suited for ROI extraction. Experiments on hyperspectral images show that the improved algorithm has not only more stable but also more accurate segmentation results.5. We concentrate on aurora ROI extraction based on FS-KFDA and frame analysis. The main work consists of two aspects.(1) Starting from aurora imaging mechanism, we study on algorithm suitable for aurora image. Aurora image has similar spatial and spectral feature to hyperspectral image, we can take advantage of the interframe correlation. We focus on feature extraction in the spectral domain. A new spectral feature exatraction algorithm is constructed on discrete wavelet transform.(2) We train the FS-KFDA classifier with spatial and spectral feature, in order to make it suitable for aurora sequence image. Experiments show that the new algorithm has better classification and is more suitable for aurora sequence image than other classifiers.