A survey of hyperspectral image super-resolution method

不同于传统图像（如灰度图像、RGB图像等）专注于保存目标场景的空间信息，高光谱图像蕴含丰富的空—谱信息，不仅可以保存目标的空间信息，还可以保存具有高可辨性的光谱信息。因此高光谱图像广泛应用于多种计算机视觉和遥感图像任务中，如目标检测、场景分类和目标追踪等。然而，在高光谱图像获取以及重建过程中仍然存在许多问题与瓶颈。如传统高光谱成像仪器在成像过程中通常会引入噪声，且获得的图像往往具有较低的空间分辨率，极大地影响了高光谱图像的质量，对后续数据分析任务造成了极大的困难。近年来，高光谱图像超分辨率重建技术研究得到了极大的发展，现有超分辨率重建方法可以大致分为两类，一类为空间超分辨率重建方法，可以通过直接提升高光谱图像的空间分辨率来获得高质量高光谱图像；另一类为光谱超分辨率重建方法，可以通过提升高空间分辨率图像的光谱分辨率来生成高质量高光谱图像。本文从高光谱图像超分辨率重建领域的新设计、新方法和应用场景出发，通过综合国内外前沿文献来梳理该领域的主要发展，重点论述高光谱图像超分辨率重建领域的发展现状、前沿动态、热点问题及趋势。;Computer vision-oriented hyperspectral images(HSIs) are featured to enrich spatial and spectral information compared gray RGB images. It has been developing in such domains like target detection, scene classification, tracking. However, the HSI imaging technique is challenged for its distortion problems(e. g., low resolution, noise). HSI-related super-resolution(SR) methods proposed reconstruct high-quality HSIs terms of a high resolution resolution(HR). Current SR can be segmented into two categories:spatial SR. The method oriented HR via improving low-resolution(LR) HSI. subdivided single image fusion based further. Single used directly LR due much more sacrificed, effective HSIs.Therefore, fuse these quality HSIs, extra homogeneous introduced(e. multispectral image(MSI), RGB). In this way, improved greatly(e. 8 times, 16 32 times SR). other focused on super-resolution method, which improve images(e. MSI, RGB) generate image. literature review growth relevance three aspects image-based, fusion-based, contexts. Each category further traditional optimization framework deep learning methods. For SR, problem being an illness inverse problem, develop priors restrain process. Image low-rank, sparse representation, non-local features commonly but it still problems manpower restrictions. core element balance spatial-spectral correlation between MSI feasible split images key components then re-combine parts each Therefore, multiple schemes nonnegative matrix factorization, coupled tensor factorization) leverage from addition, increase effectiveness decomposition methods, some constraints required introduced (e. sparse, low-rank). essential learn how characteristics RGB/MSI When paired exist, promising way construct dictionary(e. dictionary learning) that mapping relation recorded HSIs. have as well, often realize generalization ability applications. recent years, facilitated computer vision tasks, beneficial exploiting inherent relations convolution neural network(DCNN) utilized process DCNN capable prior plenty training samples, better representation than heuristic handcrafted priors(e. sparse) extent. performance kind restricted by amount variety samples. necessary unsupervised manner, robustness challenging resolved. extract MSIs design frameworks(e. multi-branch, multi-scale, 3D-CNN) concerned about. noise, unknown degeneration, unregistered) DCNN-based method. To resolve problems, alternative fusion-based ability. DCNN-unfolding optimal interpretability demonstrated, registration strategy super demonstrated model scheme there barriers resolved example, most existing only according fixed interval or spectrum range. A linked with well future. study, we will summarize scope perspective new designs, application scenes.