基于兴趣体保护的高光谱遥感图像压缩技术研究

发布时间：2019-01-03 21:47

【摘要】：较高的光谱分辨率对于高光谱遥感图像的应用十分重要,但是,高光谱图像带来了海量数据,这给数据的存储和传输带来了巨大压力。高光谱遥感图像自身特性决定其不同于二维图像,因此从高光谱遥感图像自身特性出发来进行高光谱遥感图像的压缩就是一个重要的研究课题。论文探索了高光谱遥感图像的空间相关性和谱间相关性,研究其与静态二维图像之间的相关性差异。结果表明：高光谱图像的空间相关性要弱于二维静止图像,同时高光谱图像的谱间相关性要强于空间相关性。提出了兴趣体的概念,即感兴趣波段中的感兴趣区域。针对地物分类应用,研究了最佳波段选择方法；针对目标提取应用,研究了基于小波分解的波段选取方法。并通过仿真实验验证了两种波段选择的效果。在ROI提取过程中,论文采用了FastICA算法,并通过实验对比证明了该方法在突出目标以及抑制背景两个方面都明显好于其他常见方法,主要原因是事先利用MNF变换对原始数据进行降维,从而降低了噪声对目标探测的不利影响。论文研究了基于小波Contourlet变换和兴趣体保护的高光谱图像压缩方法。可分离二维小波变换对纹理丰富图像进行低比特率压缩时,图像边缘附近会产生“振铃”现象,而单纯的Contourlet变换具有4/3冗余度,导致Contourlet系数增多,不利于图像的压缩编码,因此研究了二者结合、优势互补的图像压缩算法。兴趣体保护较单纯的兴趣区保护或兴趣波段保护而言,保护内容更加精准,需保护的信息量大大减少,因此在进行相同码率压缩情况下,重构图像的兴趣区与背景区都将获得更优的图像质量。将Maxshift算法推广到三维图像,提升了BOI中ROI区域的变换系数。通过实验验证了在压缩比8：1时,本文提出的算法峰值信噪比达到了40dB以上,较单纯使用小波变换的3D-SPIHT算法提高了3.16dB；从视觉效果可以看到由于结合了Contourlet变换,重构图像纹理清晰,边缘明显。重构图像用于分类应用几乎未受影响。
[Abstract]:High spectral resolution is very important for the application of hyperspectral remote sensing images, but hyperspectral images bring massive data, which brings great pressure to the storage and transmission of data. Hyperspectral remote sensing images are different from two-dimensional images because of their own characteristics. Therefore, it is an important research topic to compress hyperspectral remote sensing images from their own characteristics. The spatial correlation and spectral correlation of hyperspectral remote sensing images are explored in this paper, and the differences between spatial correlation and static two-dimensional images are studied. The results show that the spatial correlation of hyperspectral images is weaker than that of two-dimensional still images, and the spectral correlation of hyperspectral images is stronger than that of spatial correlation. In this paper, the concept of the part of interest is proposed, that is, the region of interest in the band of interest. The optimal band selection method is studied for the application of ground object classification, and the wavelet decomposition based band selection method is studied for the target extraction application. The effect of two bands selection is verified by simulation experiment. In the process of ROI extraction, FastICA algorithm is used in this paper, and the experimental results show that this method is better than other common methods in highlighting target and suppressing background. The main reason is that the MNF transform is used to reduce the dimension of the original data, which reduces the adverse effect of noise on target detection. In this paper, a method of hyperspectral image compression based on wavelet Contourlet transform and object of interest protection is studied. When the image is compressed at low bit rate by separable two-dimensional wavelet transform, the phenomenon of ringing will occur near the edge of the image, while the simple Contourlet transform has 4 / 3 redundancy, which leads to the increase of the Contourlet coefficient, which is not conducive to the compression and coding of the image. Therefore, an image compression algorithm based on the combination of the two and complementary advantages is studied. The object protection is more accurate than the simple protection of the area of interest or the protection of interest band, and the amount of information needed to be protected is greatly reduced, so in the case of the same bit rate compression, The region of interest and background of reconstructed image will achieve better image quality. The Maxshift algorithm is extended to 3D image, and the transform coefficient of ROI region in BOI is improved. The experimental results show that the peak SNR of the proposed algorithm is above 40dB when the compression ratio is 8:1, which is 3.16 dB higher than that of the 3D-SPIHT algorithm using wavelet transform alone. It can be seen from the visual effect that the reconstructed image has clear texture and obvious edge due to the combination of Contourlet transform. The application of reconstructed image for classification is almost unaffected.
【学位授予单位】：辽宁工程技术大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751

【相似文献】