基于纹理上下文的遥感图像目标识别

Remote sensing images target recognition based on the texture context

基于统计流形理论并结合形状上下文思想,提出了能够描述图像纹理目标的纹理上下文特征,进而实现对遥感图像纹理目标的识别。首先将图像的灰度概率密度函数看作统计流形上的点,用所得到的图像统计流形模型来描述图像的纹理上下文特征;然后使用Fisher信息距离来度量流形上点之间的相似度,并利用匈牙利算法来匹配纹理上下文特征;最后通过计算匹配距离来实现不同图像目标的相似性度量。实验表明,与经典的灰度共生矩阵、局部二值模式和统计流形算法相比,对于具有纹理特征的遥感图像,该方法具有更高的识别率且具有普适性和稳健性。

This article, based on the statistical manifold theory and combining the idea of shape context, puts forward the texture context feature which can describe the image texture target, and then realize the recognition of remote sensing image texture targets. Firstly, take the image of the gray level probability density function as the point of statistic manifold, and the statistical manifold model is then obtained to describe texture con- text characteristics of the image. Then we use Fisher information distance to measure similarity between points on the manifold, and the Hungary algorithm is used to match the texture context characteristics. Finally calcu- late matching distance to achieve similarity measure between different image targets. Experiments show that, compared with the classical gray level co-occurrence matrix, local binary patterns and statistical manifold algorithm, this method shows higher recognition for the remote sensing images with texture characteristic, and has universality and robustness.