基于卷积神经网络的高光谱图像信息恢复技术研究

发布时间：2018-05-30 19:23

本文选题：高光谱图像 + 信息恢复　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：高光谱图像的数据有着丰富的空间信息和光谱信息,能够更好的反映地物的实际情况,在民用和军事领域有着巨大的应用前景。高光谱图像容易受到外界因素的影响,出现图像质量下降,信息丢失的情况,对图像的后续处理带来了一定的问题,所以图像的预处理尤为重要。高光谱图像的信息恢复技术作为高光谱图像的预处理过程用到的技术,一直是遥感领域研究的热点问题之一。其中,条带去除和超分辨重建是高光谱图像信息恢复中两个重要的问题。由于诸多因素的影响,高光谱图像在获取和传输的过程中,容易产生条带状的噪声,使其丢失了大量的重要信息,为图像接下来的处理带来了巨大的阻碍。因此,对高光谱图像进行条带去除是其预处理中的较为重要的一步。对于条带去除而言,现有的方法在处理条带缺失列数较多、地物较为复杂时,取得的条带去除效果较差。而基于深度学习的卷积神经网络具有良好的边缘特征学习能力和挖掘海量信息背后隐藏的信息的能力,卷积神经网络作为时下较为流行的机器学习技术,在图像处理领域已有着广泛应用。本篇论文就把深度卷积神经网络运用到条带去除中来,实验表明,当缺失条带列数较多时,该方法能取得比传统方法更好的条带去除效果。高光谱图像具有较高的光谱分辨率和较低的空间分辨率,较低的空间分辨率大大限制了高光谱图像的实际应用。从硬件上来提高高光谱图像的分辨率代价较高,于是本篇论文把超分辨率重建的方法运用到提高高光谱图像的空间分辨率和光谱分辨率中来。在超分辨率重建中,基于深度卷积神经网络的方法能够实现端对端的重建,是时下较为成功的重建方法。本文针对高光谱图像的特点,构建一维、二维和三维的卷积神经网络来分别恢复高光谱图像的光谱信息、空间信息和空-谱信息。实验结果表明,基于深度卷积神经网络的超分辨率重建比传统的方法能够更好地恢复高光谱图像的光谱和空间信息,尤其是空间信息,使得该方法有着巨大的实际应用前景。
[Abstract]:The data of hyperspectral images have abundant spatial and spectral information, which can better reflect the actual situation of ground objects, and have great application prospects in civil and military fields. Hyperspectral image is easy to be affected by external factors, the image quality decline, information loss, which brings some problems to the subsequent processing of the image, so the image preprocessing is particularly important. As a preprocessing technology of hyperspectral image, information restoration technology of hyperspectral image has been one of the hot issues in remote sensing field. Strip removal and super-resolution reconstruction are two important problems in hyperspectral image information restoration. Because of the influence of many factors, the hyperspectral image is easy to produce banded noise in the process of acquisition and transmission, resulting in the loss of a large number of important information, which brings great obstacles to the next processing of the image. Therefore, the strip removal of hyperspectral images is an important step in the preprocessing of hyperspectral images. For strip removal, the existing methods are less effective when the number of missing strips is more and the features are more complex. The convolution neural network based on deep learning has good edge feature learning ability and the ability to mine the hidden information behind massive information. Convolution neural network is a popular machine learning technology. It has been widely used in the field of image processing. In this paper, the deep convolution neural network is applied to strip removal. Experiments show that this method can achieve better strip removal effect than the traditional method when the number of missing bands is more. Hyperspectral images have higher spectral resolution and lower spatial resolution, and the lower spatial resolution greatly limits the practical application of hyperspectral images. It is very expensive to improve the resolution of hyperspectral image by hardware, so this paper applies the method of super-resolution reconstruction to improve the spatial resolution and spectral resolution of hyperspectral image. In super-resolution reconstruction, the method based on deep convolution neural network can realize end-to-end reconstruction, which is a successful reconstruction method. According to the characteristics of hyperspectral images, a convolution neural network of one dimension, two dimensions and three dimensions is constructed to recover the spectral information, spatial information and space-spectrum information of hyperspectral images respectively. The experimental results show that the super-resolution reconstruction based on deep convolution neural network can recover the spectral and spatial information of hyperspectral images better than the traditional method.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】