【作者】 阿吉特·古玛勒雅巴；

【导师】 张晔；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2007， 博士

【摘要】 随着光谱成像技术的发展,人们已经可以从不同的波段上获得目标物体非常丰富的图像信息,但是随之而来的问题是所获取图像的信息量、存储量、光谱分辨率、光谱范围的急剧增加。因此,对这些大容量、高细节的遥感数据进行编码、存储、传输和访问(无损其内在形状和边缘细节)就变得非常重要,它们会直接影响机载平台(如AVIRIS)上编解码硬件系统及算法的处理能力和效率。通常情况下,遥感数据压缩主要是减少传感器或扫描仪的数据量。尤其是在20世纪70年代的图像压缩研究中,许多学者提出了对许多领域的图像(2维或3维)数据量的压缩,因此出现了很多压缩标准和算法,例如:JPEG、JPEG2000、H.26X等,其中H.26X是针对视频序列的压缩方法,而EZW,SPIHT和EBCOT是一些新的算法。那么这些研究和普通的图像压缩有什么区别呢?这些研究不仅从2维或3维图像的数据量上去压缩,而且将编码技术和有效的图像信息保存编码结合起来去压缩图像。本研究不仅包括对2维和3维图像在大小上的压缩,而且包括对于图像编码过程中如何提高效率和如何保护图像信息的关键技术需求和关键问题的研究。主要地来说,为了解决对遥感图像和视频帧信号进行高效编码中存在的问题,本文主要改进了相应的技术和算法,提出并使用了新的编码方法和可靠技术来减少当前方法的缺点和复杂度。同时,本文阐述了现有技术的缺点和三个可以用于对三类遥感图像编码的改进之处。这三类图像为:1)纹理信息少的遥感图像;2)3维的遥感图像,重点是高光谱图像;3)具有两类图像特征的丰富纹理信息图像;具有直线特征边界(RT-SE)的丰富纹理图像以及具有形状轮廓(RT-CB)的丰富纹理图像。本文中的三个相关的重要方面为:对遥感图像ROI压缩增强;对遥感图像JPEG2000压缩和新的方向性边缘增强编码;对丰富纹理信息遥感图像变换编码增强技术。对于上述问题的考虑可以解决遥感图像编码的六大需求,并且能够在图像编码的过程中保护图像的信息。这六个方面包括:(1)为了提高现有编码器的效率,可以采用ROI(感兴趣区域)编码技术,由此引进改进的2维ROI图像编码器;(2)对高光谱图像ROI选取方法的选择和编码效率需求;(3)探索现有编码技术对大区域ROI二维图像编码的基准;(4)探索在JPEG 2000中基于EBCOT图像编码算法的关键技术,解决对2维遥感图像编码过程中的3个问题,去除JPEG2000缺乏方向性边缘的缺陷;(5)为了避免使用WPT编码的复杂性,对RT-SE选择有效的小波滤波器系数;(6)解决了现有编码器不能对具有轮廓边界线丰富纹理图像信息在编码过程中的保留问题。ROI的图像压缩编码技术是图像压缩的一种方法,本文在第三章中将详细阐述如何将其应用于遥感图像处理。本文主要致力于基于SPIHT的改进算法关键技术的研究,因为SPIHT与JPEG2000相比较容易实现。本文主要解决的关键问题如下:(1)本文提出了一种效率高、简单、可信的ROI,将其用于对纹理细节少的遥感图像压缩的方法,其中采用了新的ROI思想、改进小波系数的SOT算法、编码器优化层次分解。在实现过程中,根据常用的小波滤波器结构改进了ROI编码的编码效率。对于ROI压缩应用的主观和客观评价表明,symlet-4的效果要好于双正交小波4.4和双正交小波6.8。本文设计的编码器输出性能的RD特性比JPEG2000编码器在高码率下低,而在低码率下高。(2)超光谱遥感图像的的数据量较大,且数据为3维;因此需要更先进更有效的超光谱遥感图像ROI编码技术。因此在压缩时需要采用多线程的算法选择低比特率的ROI进行预览,然后在ROI的编码通道采用基于整数小波变换(IWT)、不对称树(AT) 3-D SPIHT的改进ROI编码方法进行超光谱图像的压缩。本文提出的ROI编码方法与通用的方法相比有一定的改善。在本文的第4章中,对配置较低的遥感图像编码环境下基于EBCOT图像编码算法的JPEG2000图像编码技术进行了应用性的研究。这里有4个关键问题,最终的成果如下:1)采用了可恢复模式下的区域图像划分技术,使得编码端的内存使用量与整块划分技术相比最多可减少20%(0.75bpp);2)在解码端采用了分辨率优先排布模式对较小图像的快速预览,解码端的内存消耗减少80%(与编码端相比);3)采用每一层平均阈值的新方法,一系列遥感图像(大小相同)以相同的图像质量被重建;4)尽管JPEG2000许多强大的特性,由于其预设定的小波变换,其在对丰富纹理图像方向性滤波能力仍有所缺陷。因此,JPEG2000中的小波变换对丰富纹理的边沿信息不敏感,为此我们提出了一种新的改进的方向小波包变换(ADWPT)来增强JPEG2000的编码。ADWPT采用了基于Contourlet变换的离散滤波器组,避免了低频部分滤波的缺陷。这种编码方法仿真结果主观和客观效果评价均要好于现有的基本的编码方法。本文第5章,针对丰富纹理图像编码技术的变换做了介绍,阐述了关键问题(5)和(6)。其中第一部分,在运用小波包的前提下,提出了一种新的选择有效RT-SE图像的小波滤波器方法,解决了小波编码中的“三者相互关系”问题。这里的“三者相互关系”即为图像特征、代价函数选择的准则和小波滤波器类型之间的相互依赖关系。这种新方法对小波包利用基于熵准则而非代价函数的能量阈值方式进行基带选择,其最普遍的方式为全局能量阈值模块设定一系列的零(NZC)。在采用NZC为88%的中度损失压缩的条件下,在新方法中选取了11个滤波器中的10个进行NZC为50%的低损耗压缩,取得了较好的效果。富含边缘线条信息的遥感纹理图像需要有效地被压缩并传输,传统的小波变换和现有的变换方法在压缩具有轮廓边界线信息的遥感图像过程中效果不佳。因此,本文提出了新的Contourlet变换结合小波包(CCWPT)变换,其编解码结果从主观上评价比较好。更进一步来说,本算法的性能从客观来说比现有的基于变换的编码好,也比近期出现的基于轮廓变换方法的效果好(SNR高出0.49dB)。更多还原

【Abstract】 As imaging sensor techniques advances to acquire more information on the desired imaging objects, the acquired image volume and its spatial and spectral resolutions increase significantly. Consequently, the coding, storage, transmission and the access of such large image data sets exceed the existing processing efficiency of the remote sensing encoder systems and the capabilities of their image codec.Generally, Remote Sensing (RS) signal compression aims at reducing the size of the signals which are acquired by a sensor or a scanner. In the image coding literature, many researchers addressed the problem of 2-D or 3-D image compression in many prospective. Consequently, many coding algorithms such as SPIHT, EBCOT and the related techniques and standards were developed. Then, what makes this research’s roots different among other common image compression research?This research addresses not only the size reduction of 2-D and 3-D images but also investigates the drawbacks, complexities of the existing image coding technology and finds convenient solutions. These convenient solutions aim at efficient and information preserving coding of the three varieties of RS image categories under the three important interrelated RS areas. The three image categories are: 1) Low texture RS images, 2) Hyperspectral RS images, and 3) Rich Texture(RT) images with two kinds of image features; i) RT images with straight edges (RT-SE) and ii) RT images with contour shaped object boundaries (RT-CB). The three important RS areas deal with the enhancements of: ROI image coding, EBCOT based JPEG 2000 techniques and Transforms based RT image coding.Such consideration makes it possible to address the following six significant RS image coding issues and requirements.(1) The need of introducing an efficient, low complex and easy to modify ROI coding technique convenient for low texture 2-D RS images, (2) The need of proposing a Multitasking algorithm for time and cost efficient Hyperspectral image ROI selection and the lossless coding, (3) To explore the feasibility of currently popular baseline ROI codec, (4) To explore and to use the advanced key techniques of the EBCOT image coding algorithm based JPEG 2000 for solving the three issues deal with larger 2-D RS image coding and to eliminate the directional edge detail coding drawback of JPEG 2000, (5) The need of selecting efficient wavelet filters for RT-SE image coding while avoiding the complexities associated with the Wavelet Packet Transform(WPT), (6) The inefficiency problem of the existing transforms for information preserving coding of RT-CB remote sensing images are addressed and explained in the following paragraphs.The ROI image coding technology enhancements and research for RS are presented in Chapter 3. The key issues (1) and (2) are discussed bellow.(1) We propose an efficient, low complexity ROI image codec for compressing low texture (LT) RS images. It uses a convenient ROI coding concept, modified Spatial Orientation Trees (SOT) of wavelet coefficients, an optimal decomposition level for the SPIHT based LT image ROI codec. The set of wavelet filters used ROI codec efficiency explorations reveal that for faster RS image ROI coding tasks, the Symlet-4 adaptation performs better. The RD performance result outperformed that of the EBCOT based JPEG 2000, fairly at low bit rate. (2) The related problems of choosing and coding a ROI image efficiently without using many expensive Hyperspectral imaging missions has arisen with the major practical limitations such as massiveness of the acquired images versus limited onboard capacity, downlink contact time, etc. Therefore, the Multitasking algorithm is proposed in order to initially select the needy ROI with the low bit rate fast preview and with reference to the images in the size reduced storage archive. Then, on the ROI coding path, it addressed the drawbacks of the existing Hyperspectral image ROI coding technology using the de-coupled integer wavelet transform (IWT), Asymmetric Tree(AT) 3-D SPIHT based modified ROI coding technique. The RD performance of the proposed ROI coding technique outperformed that of the conventional 3-D SPIHT based technique.In Chapter 4, the EBCOT image coding algorithm based advanced JPEG 2000 technology is innovatively used for low resourceful RS encoder environment and the above key issue (4) is addressed. The ultimate outcomes are as follows.1) The use of the Precinct Partition image division strategy under the reversible image coding mode could achieve a maximum of 20 percent (at 0.75bpp) decrease of the encoder overall memory consumption than the use of Tiling Partition strategy under the reversible mode. 2) The 80% lowering of memory consumption at decoder and the immediate undersized on-ground preview image are achieved using the resolution prioritized progression ordering mode. 3) Using our enhanced algorithm with average layer thresholds, a sequence of different RS images is reconstructed at the same quality. 4) Despite of its many powerful features, the JPEG 2000 still lacks the RT image directional detail filtering capability due to its pre-defined wavelet transform phase. Our proposed enhancement for JPEG 2000 is named; Advanced Directional-Wavelet Packet Transform (ADWPT). The ADWPT is formed by eliminating the low frequency component filtering weakness of the Directional Filter bank Transform (DFT) associated with the Contourlet transform. The simulation results give competent performance over that of the popular, existing elementary coding techniques.Addressing the key issues (5) and (6), the transforms based coding technology enhancements for RT images are presented in Chapter 5. Initially, the new approach algorithm to select efficient wavelet filters for coding RT-SE images is formulated addressing the“Triple problem”of complexity associated with WPT based compression. This algorithm adapts the entropy-based criterion of energy threshold type as the cost function for the best basis selection. Then, a common compression for all the filters in the test set is achieved by setting the number of zeroed coefficients (NZC) as the independent parameter in the global energy thresholding module, which is a new technique. Starting with moderately lossy compression using NZC=88%, the objective performances of 11 convenient filters are observed and filter efficiency rankings are evaluated. The fairness of our new approach is validated by 10 filters out of 11 by establishing an assumption and performing the procedure of the new algorithm at NZC=50% lower lossy compression.RT-CB RS images have to be efficiently compressed and transmitted. Nevertheless, the popular wavelet transform (WT)s and the other existing transforms are less capable for the task. Therefore, in Chapter 5, the novel Contourlet Transform (CT) Combined Wavelet Packet Transform (CCWPT) is introduced; better subjective and objective coding results are obtained over the existing transforms based coding techniques. Furthermore, the SNR result outperforms that of the CT-DWT based combined technique available in the resent literature by 0.49 dB.更多还原