【作者】 苏令华；

【导师】 万建伟；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2007， 博士

【摘要】 作为一种新兴的遥感技术,高光谱遥感在军事侦察和国民经济各领域的应用日趋深入。在其应用过程中,一个瓶颈问题就是数据量过于庞大,并有愈演愈烈之势。高光谱图像压缩技术旨在解决这一难题,具有较强的理论和现实意义,因而得到了国内外学者的广泛重视。本文在总结现有压缩方案的基础上,研究了高光谱图像的压缩方案及压缩算法的性能评估问题。论文研究了基于非监督分类预处理的高光谱图像无损压缩方法。结合高光谱图像纹理细密特点,在聚类、波段重排基础上,提出了基于同类邻点预测的无损压缩方案。结合数据的空间和谱间相关性,提出了以单空间邻点、多波段像素为预测数据的压缩方案,继而将一种四选一的多预测器框架,引入到压缩处理过程,实现了较高的无损压缩比。论文研究了高光谱图像压缩算法的性能评估问题。首先对现有的高光谱图像有损压缩质量评估技术进行了分类和综述。结合光谱分类的数据应用,提出了基于最优性能的压缩性能评估方案。探讨了压缩重建图像的客观失真参数与应用准确率之间的关系,在不同地物成像实验统计基础上,提出了一种基于失真参数提取的、稳健的性能评估框架。以三种基于矢量量化有损压缩方案为例,说明了评估过程。两种评估框架都具有开放的结构,可拓展到目标检测等其它应用方向的性能评估。结合工程实践,研制了“高速数据压缩设备测试系统”,该系统设计灵活,经过简单的软硬件更改,即可用于对高光谱图像压缩设备的速度及差错控制性能的测试。论文研究了将小波变换与矢量量化相结合的有损压缩方案。在分析WKV(discrete Wavelet transform and Kronecker gain shape Vector quantization)算法的基础上,提出了分组处理的GWKV(Grouping WKV)压缩方案。该算法引入了一种等长分组或局部最优变长分组算法,实现了“2D-DWT+波段分组+KRGSVQ”的压缩流程。该压缩方案具有低复杂度、低内存需求和良好的可并行性特点。仿真实验证实了算法的有效性。论文研究了基于独立分量分析的高光谱图像压缩方案,用于特定应用条件下的压缩问题。在军事侦察中,应用集中在小目标和异常的检测,图像数据的重建不再成为必需。参考快速独立分量分析结合小波变换编码的框架VAIW(Virtualdimensionality,ATGP,fastICA,Wavelet transform),将非监督的正交子空间投影、最大距离端元提取两种几何端元提取方法,以及RX异常检测算子,引入到快速独立分量分析的混合矩阵初始化。结合端元提取,提出了一种“保守”的虚拟维数估计修正方法,解决小目标或异常易丢失的问题,实现了RVEIS-STD(RevisedVD+Endmember extraction+fastICA+SPIHT for Small Target Detection)压缩算法。仿真实验中,验证了改进的约束最小能量算子在独立分量图序列中检测目标的有效性。在某机场AVIRIS数据中,构造两个小目标,进行了小目标检测实验。实验结果证实了小目标或异常检测应用环境下RVEIS-STD压缩方案的有效性。虚拟维数修正对小目标检测的重要性同时得到了证明。论文最后对下一步的继续研究,从个人观点,进行了展望。更多还原

【Abstract】 Hyperspectral remote sensing is gainning more attention in the fields of military surveillance and national economy. However, the applications of hyperspectral image data are still in their infancy as handling the significant size of the data presents a challenge for the user community. The request for efficient compression methods becomes pressing, since technological progresses make higher and higher spatial and spectral resolutions available. In this dissertation, algorithms of hyperspectral remote sensing image compression and performance evaluation of lossy compression methods are proposed.The lossless compression mthods using unsupervised classification as pretreatment step are studied at first. Considering the detailed texture of hyperspectral image, a predicting scheme using adjacent pixels in the same cluster is presented based on clustering and bands reordering. According to the spectral and spatial correlation characteristic, another predicting scheme using single adjacent pixel and self location pixel in multi-bands as predicting data is proposed. A multi-predictor frame with one from four strategy is presented consequently. The experimental results show the validity of the proposed algorithms.In some cases, especially in satellite data link, limited by the available bandwidth and the onboard storage capacity, lossy compression techniques have to be used. The evaluation of the quality of reconstructed data becomes a new issure. In chapter 3 of this dissertation, the existent evaluation methods are classified and summarized. Combined with the spectra classifying application, an evaluation scheme called optimal performance is developed. The relationship between distortion criteria and the preciseness of applications is discussed. And a robust criteria extraction method is developed based on imaging data of various scenes. Three lossy compression approaches using vector quantization are used to show the evaluation procedure. Both of the evaluation schemes have opening framework. Combined with engineering practice, a testing system for high speed data compression equipments is presented. The system is designed flexibly, and by simple changing the hardware and software, the system can be used to test the speed and error control performance of hyperspectral data compression equipments.Bands grouping idea is introduced to the WKV (two dimensional Wavelet transform and Kronecker gain shape Vector quantization) lossy compression method in chapter 4. An equal length grouping and a locally optimized grouping are used. And a new method called GWKV (Grouping WKV) is developed. Compared with the original scheme, the new method has merits of lower complexity, lower storage space and good parallelizability. The experimental results show that the algorithm is efficient.In military surveillance, the application focuses on the detection of small and anomalous targets. In some cases, the image reconstruction is not necessary. Compression methods based on independent component analysis (ICA) are discussed in chapter 5. Two geometric endmember extraction methods: unsupervised orthogonal subspace projection (UOSP) and maximum distance extractor, and RX anomality detector,are introduced to the mixing matrix initialization of FastICA. Combined with endmember extraction, a "conservative" revising algorithm of the estimation of virtual dimensionality (VD) is proposed to satisfy the need of small targets detection. The compression method called RVEIS-STD is developed. In the experiments, the validity of the modified CEM (constrained energy minimum) operator when it is used on independent components is studied. We construct two small targets in the AVIRIS data of an airfield, and detect the small targets successfully. The results approve that the RVEIS-STD compression method is an efficient compression approach in small targets and anomalies detecting applications. Simultineously, the results also validate the key effect of the revising algorithm in small characteristics protection.In the end, the direction of further study is pointed out, though, in personal opinion.更多还原