基于非平稳表示模型的高光谱影像去噪方法研究

发布时间：2019-01-15 21:02

【摘要】：目前,高光谱遥感技术发展迅速,应用范围广泛,为了获取高质量的影像数据,在改善仪器设备有限的情况下,还可以通过影像处理的方式提高影像质量。针对获取影像数据时受到的噪声影响问题,对高光谱影像数据进行去噪处理便成为影像处理步骤中不可缺少的一步,所以本文提出了一种基于非平稳表示模型的高光谱影像去噪方法。在高光谱影像中,对于某个单独像元来说,从光谱维来看,它具有一条完整的光谱曲线,可以用于判别该像元所对应的地表内包含的地物类型;从空间维来看,可以获取该像元在地面的位置信息,及与其它像元在空间的位置上的排列组合,这是高光谱影像所独有的“图谱合一”特点。而且由于传感器的空间分辨率较低,使得每个像元所对应的地面范围较广,包含的地物较多,其像元的光谱曲线是多种地物的光谱曲线的混合。为了解决这一问题,利用高光谱影像数据独有的特点进行混合像元分解,得到“真实的”端元值和丰度值,再进行影像重建便可以得到去噪后的高光谱影像。这是基于混合光谱分解的去噪方法。在本文中所提出的非平稳表示方法是在基于混合光谱分解的去噪方法基础上更进一步的研究,主要内容有:1、根据线性光谱混合模型,将这些像素及其权重用于本征表示解混方法中,以解混的思想对中心像元进行混合像元分解,将求解得到的“真实的”端元值和丰度值进行影像重建,恢复之后的高光谱影像,即是得到的去噪后影像。2、针对高光谱影像数据具有空间非平稳性,依托由Fuentes等人提出的空间非平稳建模方法;利用非局部均值方法寻找与中心像元相似度大的像素,分别使用马氏距离和欧氏距离计算图像块之间的相似度作为权重。3、利用模拟高光谱数据和真实高光谱数据进行实验,对去噪前后的影像进行定性和定量评价。分别计算模拟影像的峰值信噪比,结构相似度,均方根误差和真实影像的信噪比。经过实验结果证明,本文模型能够很好的同时在光谱维和空间维对高光谱影像去噪,相比较其它方法,本文中所提出的方法具有很好的稳健性,能更多的保留影像的空间纹理信息。
[Abstract]:At present, hyperspectral remote sensing technology has been developed rapidly and widely used. In order to obtain high quality image data, the image quality can be improved by image processing under the condition of limited instrument and equipment. In order to solve the problem of noise impact on the acquisition of image data, de-noising of hyperspectral image data becomes an indispensable step in image processing. In this paper, a method of hyperspectral image denoising based on nonstationary representation model is proposed. In hyperspectral images, for a single pixel, it has a complete spectral curve from the perspective of spectral dimension, which can be used to distinguish the type of ground objects contained in the surface corresponding to the pixel. From the spatial dimension, the location information of the pixel on the ground and its arrangement and combination with other pixels in space can be obtained, which is the unique feature of hyperspectral images. Because of the low spatial resolution of the sensor, each pixel corresponds to a wide range of ground and contains more ground objects. The spectral curve of each pixel is a mixture of the spectral curves of a variety of ground objects. In order to solve this problem, using the unique characteristics of hyperspectral image data to decompose the mixed pixel, get the "real" end element value and abundance value, and then reconstruct the image to get the de-noised hyperspectral image. This is a denoising method based on mixed spectral decomposition. The non-stationary representation method proposed in this paper is further studied on the basis of the noise removal method based on the mixed spectral decomposition. The main contents are as follows: 1. According to the linear spectral mixing model, These pixels and their weights are used in the intrinsic representation demultiplexing method, the central pixel is decomposed with the idea of unmixing, the "real" end element value and the abundance value are reconstructed, and the hyperspectral image is restored. That is, the de-noised image. 2. Aiming at the spatial nonstationarity of hyperspectral image data, this paper relies on the spatial non-stationary modeling method proposed by Fuentes et al. The nonlocal mean method is used to find pixels with high similarity to the center pixel, and the similarity between the blocks is calculated by Markov distance and Euclidean distance respectively. 3. Experiments are carried out using simulated hyperspectral data and real hyperspectral data. The images before and after denoising were evaluated qualitatively and quantitatively. The peak signal-to-noise ratio (PSNR), structural similarity, root mean square error (RMS) and signal-to-noise ratio (SNR) of real images are calculated respectively. The experimental results show that the proposed method is robust to the hyperspectral image denoising at the same time in spectral dimension and spatial dimension. Compared with other methods, the proposed method in this paper is robust. Can retain more spatial texture information of the image.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P237

【参考文献】