光谱去相关技术在高光谱图像小波压缩中的应用

Special Decorrelation Technique Used for DWT-Based Hyperspectral Image Compression

随着数据量的不断增长,如何有效压缩高光谱图像成为影响其普及应用的一个关键问题。近年来,小波压缩技术已经被证明是高光谱图像压缩方法中很有发展前景的一个,但由于其对高光谱图像特性的利用较为有限而使其性能的进一步提升受到了限制。文章根据高光谱图像的光谱特征,提出了一种基于光谱去相关的高光谱图像小波压缩方法,设计了分块预测方法来同时去除光谱间相关性和空间相关性,并将其应用于小波压缩方法之中。首先,将高光谱图像分为几个具有高谱间相关性的图像块。然后推导出各块中波段的近似成比例的特性,并在各块分别进行基于这一特性和超光谱图像其他特性设计波段预测编码。最后,将预测用的参考波段和预测后获得的偏差数据,通过小波编码技术进行压缩。实验结果表明,所设计的方法与目前先进的超光谱压缩技术相比其性能有显著的提升。与AT-3DSPIHT算法比较,最高PSNR或SNR提升幅度均能达到4.2dB左右。此外,此方法在低比特率下的优势也十分突出。

Hyperspectral images are massive data consisting of hundreds of spectral bands and have been used in a large number of applications.With growth of spectral resolution and spatial resolution of hyperspectral data,data size increases rapidly.How to effectively compress hyperspectral image becomes a key problem that affects the development and popularization of hyperspectral image.Recently,DWT-based methods have been proved promising for hyperspectral image.But their performances are restricted because it is difficult for them to efficiently take advantage of the various properties of hyperspectral image.For the traditional wavelet transform,the specific properties of hyperspectral images are basically utilized by corresponding to characteristics of wavelet coefficients.So the present paper proposes a new DWT-based method using decorrelation technique according to the spectral characters of hyperspectral image.Block predictive coding is designed to remove the spectral correlation as well as spatial correlation simultaneously and is applied into the DWT-based method.Firstly,hyperspectral image is divided into several image blocks.The bands in a single block possess high spectral correlation.Afterwards,it is deduced that bands of a single block tend to be proportional in altitudes.Bands prediction,which is done in the range of each block respectively,is designed according to this and others characteristics of hyperspectral images.Finally,reference bands of block prediction and the deviation data obtained after block prediction are compressed by 2D-DWT algorithm and 3D-DWT algorithm respectively.Experiment results indicate that compared with the well known techniques the proposed method can significantly improve SNR and PSNR performance,even to 4.2 dB （compared with AT-3DSPIHT algorithm）.And the code efficiency at low bit rates is also competitive.