基于相关向量机的高光谱图像噪声评估算法

为了更准确的估计高光谱图像噪声强度,提出了一种基于相关向量机(RVM)的高光谱图像噪声评估算法。对该算法所采用的RVM 回归原理、残差与噪声的关系等进行了研究。首先,介绍了高光谱图像噪声评估中应用较为广泛的空间/光谱维去相关法的特点及不足。接着,对可有效进行非线性回归分析的RVM 进行了介绍。然后,针对传统的空间/光谱维去相关法在系统中存在较强的非线性关系时,得到的残差将会过大这一问题,提出利用RVM 回归分析去除具有高相关性的信号,利用得到的残差图像对噪声进行估算,从而提高评估系统的稳定性。实验结果表明:噪声强度估计精度优于8%;相比传统算法更有效。总体看,该算法可以满足自动高光谱图像噪声评估的稳定可靠、精度高等要求。

Noise estimation algorithm based on relevance vector machine for hyperspectral imagery

In order to more accurately estimate noise intensity for hyperspectral imagery, the paper proposed a noise estimation algorithm based on relevance vector machine(RVM)for hyperspectral imagery. And the algorithm that used RVM regression, residuals and noise was studied. First of all, this paper introduced the characteristics and shortage of spatial/spectral dimension decorrelation in noise estimation that used widely nowadays for hyperspectral imagery. Then, the nonlinear regression analysis of RVM was introduced. And the residuals will be too large, when there was a strong nonlinear correlation in the system for spatial/spectral dimension decorrelation. To this problem, the paper proposed a new method that used RVM regression to remove strong signal correlation and used the residual images to estimate the noise, so as to improve the stability of the assessment system. Experimental results indicate that the precision of the noise intensity is better than 8%, and show that the method is more effective compared to the traditional method. It concludes that the RVM can satisfy the system requirements of higher precision and stabilization in noise estimation for automatic hyperspectral imagery.