基于线性混合模型的高光谱图像压缩感知研究

发布时间：2018-03-09 10:07

本文选题：高光谱图像　切入点：压缩感知　出处：《西北工业大学》2015年博士论文　论文类型：学位论文

【摘要】：随着高光谱图像分辨率的不断提高,成像光谱仪获取的高光谱数据急剧增加,海量的高光谱数据给机载或星载遥感成像系统的数据存储、传输和处理带来巨大压力。传统的高光谱遥感成像面临高速率采样、海量数据的存储和传输等难以突破的难题。压缩感知理论作为一种新颖的数据采集理论将数据的采样和压缩过程巧妙地结合起来,实现采样率低于传统奈奎斯特率的数据采集和少量观测数据的精确重构,降低了对传感器的要求,有效避免了追求高分辨率带来的软硬件成本问题。压缩感知理论的诞生为解决传统高光谱遥感成像面临的瓶颈问题提供了有效的解决途径,将压缩感知技术应用于高光谱遥感成像的研究已成为当前高光谱数据采集研究的热点。本文以高光谱数据的压缩采样模式和图像的重建算法为研究对象,以高光谱图像的线性混合模型为契机,致力于寻求高效的高光谱压缩采样方式和快速、高精度的压缩感知重建算法,并针对当前成熟的推扫式和摆扫式高光谱数据采集模式研究了压缩感知成像的实现方案。论文的主要工作及取得的创新性成果如下:1.针对高光谱图像空间或光谱维压缩采样形式单一、采样效率低的问题,在分析研究空间压缩采样域数据特性的基础上,提出了一种高光谱图像的空谱压缩采样方案。高光谱数据采样先进行空间压缩采样,再对压缩后的数据进行光谱维压缩采样。实验结果表明,空谱压缩采样有助于提高采样效率,改善重构质量。2.针对传统高光谱压缩感知直接重构原始图像数据量大的问题,在光谱维压缩采样数据的重构过程中应用线性混合模型,提出了已知端元谱和未知端元谱的压缩感知重构算法,以及基于线性光谱库混合模型的压缩感知重构算法。应用线性混合模型将高光谱图像分离成丰度系数和端元特征两个小数据量的子集,利用混合像元分解算法估计丰度,稀疏优化算法提取端元,再通过估计的丰度和提取的端元来合成原图像。实验结果表明所提的重构算法由于估计的两子集数据量小,运行速度上有较大幅度的提升,同时能获得更高的峰值信噪比和更好的重建效果。3.针对高光谱图像压缩感知重构精度低、计算复杂度高的问题,提出了基于波段的分布式压缩感知和基于像元的分布式压缩感知方案。采样时将高光谱图像分成关键波段图像与压缩感知波段图像或者关键像元与压缩感知像元,并对不同类别数据采用不同的采样方式。重构时利用线性混合模型分离不同类别的观测数据,进行端元提取和丰度估计,线性谱解混算法的应用使得欠定问题的优化转化成超定方程的求解,大幅提高了重构算法的运行速度和精度。在基于波段的分布式压缩采样基础上,提出基于迭代预测的分布式压缩感知重构算法,进一步提高了高光谱压缩感知的重构精度。多个高光谱数据集的实验结果表明,所提的两种分布式压缩感知方案重构的平均信噪比远远高于压缩投影主成分分析算法和三维压缩采样算法,且重构速度比三维压缩采样算法有数量级的提升。4.针对高光谱图像压缩测量矩阵难以光学实现的问题,构造了实现空谱联合压缩采样的测量矩阵,提出了高光谱图像的空谱联合压缩感知方案。通过分析空间压缩采样的丰度系数的数据特性,构造了便于光学实现的、具有特殊结构的空间测量矩阵。重构时采用线性混合模型通过提取端元和估计丰度恢复原图像。实验结果表明,高光谱图像采用空谱联合的压缩采样,重构时利用线性混合模型进行分离重构,能实现快速、高精度的数据采集与重建的目标。5.基于摆扫式和推扫式高光谱遥感数据采集模式,研究了压缩感知成像的实现方案,设计了实现摆扫式基于像元的分布式压缩采样成像系统模型、推扫式光谱维和空谱联合压缩采样成像系统模型。采用棱镜、柱面透镜、DMD等光电器件,能使采集的数据量大幅减少而不需要进行其它压缩处理,既节省了存储空间和传输资源,又降低了数据压缩算法带来的功耗,有利于机载或星载高光谱成像的实现。仿真所设计成像系统模型压缩观测数据,采用本文所设计的重构算法进行重构,虽然重构性能稍有下降,但仍能高精度的恢复原数据。
[Abstract]:With the continuous improvement of image resolution hyperspectral imaging spectrometer, hyperspectral data acquisition has increased dramatically, the hyperspectral data for massive airborne or satellite remote sensing imaging system data storage, transmission and processing of huge pressure. The traditional hyperspectral remote sensing imaging at high sampling rate, the problem of massive data storage and transmission is difficult to break the compressed sensing theory. As a kind of novel data acquisition theory data sampling and compression process combined skillfully, achieve the sampling rate is lower than the accurate reconstruction of data acquisition and a small amount of observation data of the traditional Nyquist rate, reduce the requirement of sensor, effectively avoid the pursuit of cost of software and hardware problems brought by high resolution. Compressed sensing theory has provided the effective way to solve the bottleneck problem of traditional hyperspectral remote sensing imaging, the compressed sensing technology Study on technique applied to hyperspectral remote sensing imaging has become a hotspot of current research. The hyperspectral data acquisition based on compression sampling mode and image reconstruction algorithm for hyperspectral data as the research object, using linear mixed models of hyperspectral image as an opportunity to seek efficient compression of hyperspectral sampling and fast, compressed sensing the reconstruction algorithm with high precision, and in view of the current mature pushbroom scanning and high spectral data acquisition mode of implementation of compressed sensing imaging. The main work and innovative results are as follows: 1. for hyperspectral image space or spectral compression sampling in single form, the problem of low efficiency of sampling. Based on the analysis of spatial domain data compression sampling characteristics, presents a high spectral image space spectral compression sampling scheme. The hyperspectral data sampling to space compression Sampling of compressed data in spectral compression sampling. The experimental results show that the spectrum of air compressed sampling is helpful to improve the sampling efficiency, improve the quality of reconstructed.2. in traditional hyperspectral compressed sensing to reconstruct the original image directly to large amount of data compression using linear mixed model reconstruction process of sampling data in the spectral dimension put forward, known and unknown endmember spectral element end of compressed sensing reconstruction algorithm in spectrum, and compressed sensing reconstruction algorithm of linear spectral library based on the hybrid model. The separation of application of linear mixed model of hyperspectral image into subsets abundance coefficient and endmember features two small amount of data, using mixed pixel decomposition algorithm to estimate the abundance. The optimization algorithm of endmember extraction by sparse, endmember abundance and extraction estimation to the synthesis of the original image. Experimental results show that the proposed reconstruction algorithm due to estimation of the two data sets Small, greatly enhance the speed, also can get higher PSNR and better reconstruction effect of.3. for hyperspectral image compressed sensing reconstruction precision is low, the computational complexity of the problem, put forward the distributed compressed sensing band and pixel distributed compressed sensing scheme based on the sampling. Hyperspectral image is divided into key band image and compressed sensing image band or key pixel and compressed sensing pixels, and using different sampling methods for different categories of data. Data reconstruction using linear mixed model from different categories, endmember extraction and abundance estimation, linear spectral unmixing algorithm makes the optimal solution using the underdetermined problem into solving over determined equations, greatly improve the reconstruction speed and accuracy. Based on the band distributed compressed sampling based on the proposed base In the iterative prediction of distributed compressed sensing reconstruction algorithm, further improve the reconstruction accuracy of hyperspectral compressed sensing. A plurality of hyperspectral data sets. The experimental results show that the average signal reconstruction of two kinds of distributed compressed sensing scheme of the proposed noise ratio is much higher than the compression projection principal component analysis algorithm and three-dimensional reconstruction and compression sampling algorithm. Faster than the 3D compressive sampling algorithm of magnitude increase.4. for hyperspectral image compression measurement matrix is difficult to optical implementation problems, constructed to realize spatial spectral joint compression sampling measurement matrix, the hyperspectral image compressed sensing scheme combined with spatial spectral data characteristics. Through the analysis of space compression abundance coefficient sampling, construct for optical implementation, space measurement matrix with special structure. The reconstruction using a linear mixed model by extracting and recovering the original image endmember abundance estimation Like. The experimental results show that the hyperspectral image compression using spatial spectral joint sampling, reconstruction using linear mixed model for separation of reconstruction, to achieve rapid,.5. data acquisition and reconstruction of high precision sweeping and pushbroom hyperspectral remote sensing data acquisition mode based on the research, the realization scheme of compressed sensing imaging that is designed to achieve the scanning type pixel distributed compressed sampling imaging system based on the model of pushbroom spectral and spatial spectral compression joint sampling imaging system model. Using the cylindrical lens, prism, DMD and other optoelectronic devices, can make the collection of data without the need for a substantial reduction in the amount of other compression processing, can save storage space and transmission resources, and reduce the power data compression algorithm, is beneficial to the realization of airborne or spaceborne hyperspectral imaging. The imaging system design model using the data compression. The reconstructed algorithm is reconstructed, although the reconfiguration performance is slightly reduced, it can still restore the original data with high precision.

【学位授予单位】：西北工业大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP751

【相似文献】