自适应优化稀疏表示的遥感图像压缩重构研究

发布时间：2018-05-31 09:13

本文选题：压缩感知 + 过完备字典　；参考：《浙江大学》2014年硕士论文

【摘要】：目前,国内外各类光学遥感采样成像系统均基于奈奎斯特—香农采样理论,它指出采样率必须达到信号带宽的两倍以上才能精确地重构信号。随着遥感图像空间分辨率的提高,要求光学系统的焦距更长、口径更大,焦平面器件的采样率更高,像元面积更小,这将大大增加光学系统、焦平面器件的设计和制造难度。而压缩感知理论指出,只要信号是稀疏的或者在某一变换空间是稀疏的、可压缩的,就可以远低于奈奎斯特采样定理所规定的采样量得到信号的压缩表示,并且仍能够精确地重构原始信号。因此,在遥感成像系统中采用压缩感知理论进行图像的压缩采样,可以在采样的同时实现压缩,获取图像迅速,节约工作时间；仅少量采样值即可重构原始高分辨图像,极大地节约焦平面阵列器件,同时节省星上存储空间；而针对不同类型的遥感图像自适应地选择最优稀疏表示方法,可以有效地提高遥感图像重构质量,便于后期的信息提取工作；无论被观测图像类型如何,采用固定的观测方式,均能获得高质量重构图像。论文首先介绍了课题背景和意义,总结了压缩感知理论框架中的常见图像稀疏表示方法,图像的随机观测手段及其观测矩阵的构造,介绍了压缩感知中常见的优化重构算法,将压缩感知理论在图像复原和图像融合领域提出了自己的创新和改进方法。论文同时也介绍了几种常用的图像复原的方法,从主观和客观方面讨论了图像质量评价的作用和意义。重点介绍了采用K-SVD方法进行过完备字典训练的方法。对不同类型遥感图像的训练字典对于不同类型图像的稀疏表示性能进行了深入分析。介绍了通过给定的训练字典对随机观测矩阵进行迭代优化的联合优化方法,并且采用该方法进行实验仿真,获得了优化后的观测矩阵。同时,采用优化后的观测矩阵与训练字典对,测试了图像的重构效果。在此基础上,深入分析并总结了重构稀疏度与训练稀疏度倍率关系对于图像重构的影响,并提出了从观测值中随机选取对应于原始图像块的少量观测值序列进行小代价重构从而预估出获得良好重构质量的重构稀疏度的方法,有利于快速且准确地找到最优的重构稀疏度从而获得高重构质量。论文最后提出了自适应优化稀疏表示的遥感图像压缩重构框架,重点围绕城市、山地和海港这三类遥感图像进行了对应类型图像的字典训练,获得这三类图像的优化稀疏表示字典,并分析对比了采用固定稀疏基DCT的稀疏表示精度。通过遥感图像的压缩感知粗复原初步判别遥感图像类型,采用其对应的优化稀疏表示字典作为稀疏表示方法,再采用精确重构算法对压缩采样值进行精确重构。对重构图像采用主观评价为辅,客观评价中的PSNR和SSIM评价值进行分析。结论表明,采用本文提出的方法在提高遥感图像压缩重构质量上取得了较好的效果。
[Abstract]:At present, all kinds of optical remote sensing sampling and imaging systems at home and abroad are based on Nyquist Shannon sampling theory. It points out that the sampling rate must be more than twice of the signal bandwidth in order to accurately reconstruct the signal. With the improvement of spatial resolution of remote sensing image, the optical system is required to have longer focal length, larger aperture, higher sampling rate of focal plane device and smaller pixel area, which will greatly increase the difficulty of design and manufacture of optical system and focal plane device. Compression sensing theory states that as long as the signal is sparse or sparse and compressible in a transformation space, the compressed representation of the signal can be obtained far less than the sample amount specified by Nyquist sampling theorem. And still can accurately reconstruct the original signal. Therefore, compression sensing theory can be used to compress images in remote sensing imaging system, which can compress images at the same time, obtain images quickly and save working time. Only a small number of sampling values can reconstruct original high-resolution images. The focal plane array devices are greatly saved and the on-board storage space is saved, and the quality of remote sensing image reconstruction can be effectively improved by adaptively selecting the optimal sparse representation method for different types of remote sensing images. It is convenient for information extraction in the later period, and high quality reconstructed images can be obtained by using a fixed observation method, regardless of the type of the observed image. Firstly, the paper introduces the background and significance of the subject, summarizes the common image sparse representation methods in the theory framework of compressed perception, the random observation method of image and the construction of observation matrix, and introduces the common optimization reconstruction algorithms in compressed perception. The compression perception theory is proposed in the field of image restoration and image fusion. At the same time, several commonly used methods of image restoration are introduced, and the function and significance of image quality evaluation are discussed from subjective and objective aspects. The method of complete dictionary training using K-SVD method is introduced emphatically. The sparse representation performance of different types of remote sensing images is analyzed in detail by training dictionaries of different types of remote sensing images. A joint optimization method for iterative optimization of random observation matrix by a given training dictionary is introduced. The optimized observation matrix is obtained by using this method for experimental simulation. At the same time, the image reconstruction effect is tested by the optimized observation matrix and training dictionary. On this basis, the effect of the relationship between the sparse degree of reconstruction and the ratio of training sparsity on image reconstruction is analyzed and summarized. A method is proposed to estimate the reconstruction sparsity with good reconstruction quality by randomly selecting a small number of observation value sequences corresponding to the original image blocks from the observed values for small cost reconstruction. It is helpful to find the optimal reconstruction sparsity quickly and accurately and obtain high reconstruction quality. At the end of this paper, an adaptive and sparse image compression and reconstruction framework is proposed, which focuses on the dictionary training of the corresponding types of remote sensing images around the cities, mountains and seaports. The optimized sparse representation dictionaries of these three kinds of images are obtained, and the sparse representation accuracy using fixed sparse base DCT is analyzed and compared. The types of remote sensing images are preliminarily identified by the compressed perception rough restoration of remote sensing images. The corresponding optimized sparse representation dictionary is used as the sparse representation method, and the precise reconstruction algorithm is used to reconstruct the compressed sampling values accurately. The subjective evaluation is used as the assistant to reconstruct the image, and the PSNR and SSIM evaluation values in the objective evaluation are analyzed. The results show that the proposed method can improve the quality of remote sensing image compression and reconstruction.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】