【作者】 王朗；

【导师】 郭树旭；

【作者基本信息】 吉林大学 ， 电路与系统， 2009， 博士

【摘要】 超光谱遥感技术近些年发展迅速并得到广泛应用。但随着超光谱图像光谱和空间分辨率的增加造成的海量数据给传输和存储带来了巨大的困难。所以研究出有效的超光谱图像采样和压缩方法具有重要的理论意义和实用价值。因此,本文结合超光谱图像自身的光谱特点,提出了超光谱图像压缩的一些新方法。首先,论文从谱间的结构相关和统计相关角度出发,结合数据分析了超光谱图像具有强烈的谱间相关性,并远大于空间相关性。同时,发现超光谱图像的空间相关性也低于普通图像的空间相关性。发掘了它与多光谱图像的不同的局部非平稳特征。确定了压缩重点在于利用超光谱图像的谱间相关性以及局部不平稳特性。超光谱图像还具有高数据维的特点,通过计算得到高维谱空间大部分为空的结论,为选择向低维投影采样压缩方法提供了理论依据。本文将嵌入式编码应用于超光谱图像压缩,对SPIHT算法进行改进,在经典SPIHT算法中利用小波系数兄弟节点相关性,修改了零树结构,引入LZC算法的标志位图思想,图像恢复质量明显改善。基于超光谱图像特性,本文首先将基于侧四邻域定义的误差补偿预测树算法应用到超光谱图像无损压缩,建立了自适应双向波段预测的误差补偿预测树模型。针对模型预测器系数计算复杂度高的情况,设计了基于“权重”思想,利用已编码像素对系数完成自适应驱动估计。最后结合改进的SPIHT算法进行编码。实验表明,该算法在较低的计算复杂度下,压缩比优于目前流行的无损压缩算法。最后,将最近出现信号处理领域的新理论“压缩传感理论”应用于超光谱图像压缩。本文针对压缩传感理论中稀疏分解的计算量巨大,提出了对传统匹配跟踪算法的改进,使用改进的遗传算法来寻找基函数库最优原子,同时对图像进行分块匹配搜索,减小最优匹配搜索范围,并自适应决定稀疏分解结束。在实现了图像信号准确重构的基础上,以较低的计算复杂度完成了直接信息采样。本文围绕着超光谱图像所特有的光谱图像特性,提出并改进不同的压缩算法,并分别进行了性能分析和实验评价。更多还原

【Abstract】 Hyperspectral remote sensing based on the research of cross-discipline including the electromagnetic spectrum, geographic information systems, electronic technology, computer technology, aerospace technology and other subjects, emerges as a novel remote sensing technology and develops rapidly in recent years. Spectral resolution of the hyperspectral remote sensing, which is higher than 1% of wavelength, hits nanometer (nm) order of magnitude, and the number of spectral channels reaches up to tens or even hundreds. It organically combines the ground object spectra determined by the material composition and the space imaging inflecting the existed pattern of ground object and each pixel of space imaging can be assigned its own information of characteristic spectrum, which may improve greatly cognitive ability of the objective world. Hence, hyperspectral remote sensing has great value of application in various areas, such as survey of land resources, forestry remote sensing, environmental monitoring, agricultural applications, the space environment, and military target detection.However, with the spectral resolution of the hyperspectral ascended, the image that contains a wealth of information of ground object and spectra brings out mass data, which leads to engender hosts of difficulties in the transmission, storage and processing of image and restricts the pace of progress in the application of hyperspectral remote sensing. In other words, how to fully utilize the potential advantages of hyperspectral images and proposing effective sampling and compression methods of hyperspectral images have important theoretical significance and practical value. Therefore, this paper, which is based on the research of the spectral image compression technology of predecessors and combined with the characteristics of hyperspectral images, proposes some novel approaches of hyperspectral image compression, and these approaches perform well.Firstly, this paper makes a detailed analysis of the hyperspectral image characteristics as the basis for follow-up compression methods. More precisely, hyperspectral image has hundreds of spectral bands and a strong correlation between spectral in terms of the common image, beside, it can be found from the data that the correlation between spectral is much larger than the spatial correlation of image. In terms of the spatial correlation, the spectral image is apparently lower than the common image due to the large coverage area of ground object. Moreover, in the analysis of correlation, it can be seen that the spectral image itself has the local non-stationary characteristics, which is obvious different from the multispectral image. Consequently, these compression approaches that are applied to common image and multispectral image are not fully suitable for the hyperspectral image. In order to obtain the idea rate of compression, correlation between spectral and local non-stationary characteristics of hyperspectral image should be sufficiently used. The hyperspectral image also has the character of high-dimensional data, and conclusion that the high-dimensional spectral space is empty can be obtained by calculation. This provides a theoretical basis for the low-dimensional projection sampling compression method.Secondly, an embedded code has been applied on the compression of hyperspectral image and the design method with SPIHT (Set Partitioning in Hierarchical Trees) algorithm based on wavelet transform has been proposed. This method for compression makes efforts on performance and shortening time for coding and decoding. But the quality of image reconstruction using the method under low bit rate is not good. To solve this problem, the classic SPIHT algorithm is improved. The correlation hypothesis about neighbor nodes has been added to SPIHT algorithm and zero- tree’s structure is also modified. In the meanwhile, it broaches the mind of LZC algorithm’s symbol flag idea. Finally, the performance on hyperspectral image compression is improved.Thirdly, to improve the real-time performance of the current compression algorithms on hyperspectral image, a new lossless compression method based on prediction tree with error variances compensated for hyperspectral image is proposed in this paper. The method incorporates prediction tree and adaptive interband prediction techniques, and bidirectional interband prediction to current band is firstly applied to hyperspectral image compression. Then the error created by prediction tree is compensated by linear adaptive predictor which de-correlates spectral statistic redundancy. In consideration of the complexity for the coefficients’ calculation, a correlation-driven adaptive estimator is designed to coefficients whose parameters are uniquely determined by the previously coded pixels. After de-correlating intraband and interband redundancy, an efficient wavelet coding method, SPIHT, is used to encode residual image. Finally, the proposed method in this paper achieves both low overhead and high compression ratio on data from the NASA JPL AVIRIS than current compression methods. At last, a novel theory of information acquisition-"compressive sampling" has been applied in this paper. The proposed approach offers a different perspective with regards to common wisdom in data acquisition of Shannon theorem. Common perception in compressed sensing indicates that one can recover certain signals and images perfectly from far fewer samples or measurements than traditional methods use. This paper presents an improvement on genetic algorithm instead of match pursuit algorithm in consideration of the enormous computational complexity on sparse decomposition. When applied to image data, our proposed approach divides the original scene into small blocks which can be processed by sparse decomposition, and a stopping criterion to decomposition process is determined by a peak signal to noise ratio (PSNR) threshold in adaptive fashion. Our experimental results indicate that good performance on image reconstruction with less computational complexity can be achieved.This paper centers on particular spectral feature of hypersectal image. Different compression algorithms proposed have been improved for need. The performance analyses and experiment evaluation are respectively given.更多还原