空间—谱间字典的学习及基于字典的高光谱图像的重构

发布时间：2018-06-12 06:02

本文选题：高光谱图像 + 字典学习　；参考：《河北大学》2014年硕士论文

【摘要】：高光谱图像的光谱具有显著的结构特征，如果高光谱图像得到适当的表征可以实现更高效的数据采集并且能够提高数据的分析能力。因为大部分像素所反映的只是少数的几种材料光谱反射曲线，因此我们认为稀疏编码模型与高光谱图像数据是良好匹配的。稀疏模型认为每个像素只是一个较大的字典中几个元素的组合，并且这种方法在应用中被证明很有效。本文提出了一种新的空间-谱间字典的学习方法，并用这个字典进行高光谱图像的重构。本文采用梯度下降法学习字典，并对梯度下降法做了简要的介绍。同时，本文提出了空间-谱间字典的学习基本思路。首先，，初始化字典取随机正值,固定字典利用梯度下降法计算稀疏系数；其次，系数不变再用梯度下降法训练更新字典；最后，上述两步交替进行直到算法收敛。依据这种模型训练出来的字典更加符合高光谱图像的特点，并将训练出来的字典用于高光谱图像的重构，通过比较峰值信噪比PSNR来确定图像重构效果的好坏，本文通过字典重构的图像的PSNR与原始图像比较获得了良好的重构效果。
[Abstract]:The spectrum of hyperspectral image has significant structural features , and if the hyperspectral image is properly characterized , more efficient data acquisition can be realized and the analytical capability of data can be improved . Because most of the pixels reflect only a small number of spectral reflection curves of the material , we think that the sparse coding model is well matched with hyperspectral image data . The sparse model considers that each pixel is a combination of several elements in a larger dictionary , and this method is proved to be effective in application .

This paper presents a new learning method of space - spectrum dictionary , and uses this dictionary to reconstruct hyperspectral image . In this paper , the gradient descent method is used to study the dictionary , and the gradient descent method is introduced briefly . At the same time , this paper presents the basic idea of learning the space - spectrum dictionary . First , the initialization dictionary takes the random value , and the fixed dictionary calculates the sparse coefficient by gradient descent method .
secondly , training the updating dictionary by a gradient descent method ;
finally , the two steps are alternately carried out until the algorithm converges , the dictionary trained by the model is more consistent with the characteristics of hyperspectral images , and the trained dictionary is used for reconstruction of hyperspectral images , and the quality of the image reconstruction effect is determined by comparing the peak signal - to - noise ratio psnr , and a good reconstruction effect is obtained by comparing the psnr of the image reconstructed by the dictionary and the original image .
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】