基于双通道的压缩光谱成像及其重构算法GPU实现

发布时间：2018-03-01 11:00

本文关键词： 光谱成像压缩感知结构稀疏聚类互补双通道 GPU　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：光谱成像技术在军事侦察、农业生产、医疗诊断、科学研究等领域意义重大，然而传统的光谱图像获取方法由于在时间分辨率、数据传输、系统开销等方面存在较大压力，因此难以满足各领域的需求。孔径快照光谱成像仪（Coded ApertureSnapshot Spectral Imagers, CASSI）是一种新型的基于压缩感知理论的光谱成像系统（简称：压缩光谱成像）。它可以通过一次曝光实现对一帧光谱场景的随机混叠观测，极大的降低了采样速率、数据传输等各方面的压力。但由于它采用的是单通道的压缩观测模型，损失了大量的光谱信息，很大程度上降低了信噪比，因此光谱图像重构质量不尽如人意。针对该问题，本文创新性地提出了基于互补编码模板的双通道光谱观测模型，该模型充分利用每一帧场景的光谱能量和信息，提高了信噪比，实现了对光谱场景的互补观测，使得光谱图像重构质量得到较大提升。另一方面，现有的光谱图像重构算法普遍采用固定变换空间作为稀疏域，难以很好地描述光谱图像中的非局部结构相似性，也未能利用光谱图像的谱间相关性信息。针对这一问题，本文利用图像的局部自相似性和非局部的自相似性提出针对光谱图像的基于结构稀疏聚类的表示模型。仿真实验结果表明本文算法能获得高质量的光谱重构效果。基于结构稀疏聚类的光谱重构算法内部包含多个子算法，计算复杂度较高，不利于光谱图像重建的实时性。为此，本文又设计了基于GPU的并行计算方案，测试实验表明了本文方案加速效果明显，，且搭建系统框架简单。
[Abstract]:Spectral imaging technology is of great significance in the fields of military reconnaissance, agricultural production, medical diagnosis, scientific research and so on. However, the traditional spectral image acquisition methods are under great pressure in terms of time resolution, data transmission, system overhead, etc. Therefore, it is difficult to meet the needs of various fields. The aperture snapshot spectral imager (CASSI) is a new type of spectral imaging system based on compression sensing theory (abbreviated as compressed spectral imaging). It can be realized by one exposure. Random aliasing observation of a spectral scene, The pressure of sampling rate and data transmission is greatly reduced. However, because it uses a single channel compression observation model, it loses a lot of spectral information, and greatly reduces the signal-to-noise ratio (SNR). Therefore, the quality of spectral image reconstruction is unsatisfactory. To solve this problem, a novel dual-channel spectral observation model based on complementary coding template is proposed in this paper. The model makes full use of the spectral energy and information of each frame scene, improves the signal-to-noise ratio, and realizes the complementary observation of the spectral scene. The quality of spectral image reconstruction is greatly improved. On the other hand, the existing spectral image reconstruction algorithms generally use fixed transform space as sparse domain, so it is difficult to describe the similarity of non-local structure in spectral image. The spectral correlation information of spectral images is also not used. In response to this problem, In this paper, a representation model based on structural sparse clustering for spectral images is proposed by using local self-similarity and non-local self-similarity of images. The simulation results show that the proposed algorithm can achieve high quality spectral reconstruction. The spectral reconstruction algorithm based on structural sparse clustering contains multiple subalgorithms, which has high computational complexity and is not conducive to real-time spectral image reconstruction. Therefore, a parallel computing scheme based on GPU is designed in this paper. The test results show that the acceleration effect of the scheme is obvious and the system framework is simple.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751;O433

【参考文献】