一种基于多层次MRF的多源图象融合算法

A Multisource Image Fusion Method Based on Hierarchical Markov Random Field Model

图象融合技术的主要目的是将多种图象传感器数据中的互补信息组合起来 ,使形成的新图象更适合于计算机处理 (如分割、特征提取和目标识别 )等 .在多层次 MRF模型的基础上 ,提出了一种应用于多源图象分类的图象融合算法 .该融合算法将定义在多层次图结构上的非线性因果 Markov模型与贝叶斯 SMAP(sequential m axi-mum a posteriori)最优化准则结合起来 ,克服了 MAP(maximum a posteriori)准则在多层次图结构上计算不合理的缺陷 .该算法可应用于多源遥感图象中的信息融合 ,使像素分类更精确 ,并解决多源海量数据的富集表示 .另外还利用合成图象与自然图象分别针对多层次 MRF模型的改进及算法中可最优化准则的不同进行了对比实验 ,结果表明 。

The main objective of the technique of image data fusion is combining or amalgamating information from multiple sensors such that the new images are more suitable for the purpose of the computer processing tasks such as segmentation, feature extraction, and object recognition. This paper proposes a multisource image fusion algorithm based on hierarchical markov random field model. The algorithm defines the fused image as the hidden labels, and the multisource image can be defined as observation of the hidden labels. The hidden labels can be extended in the quadtree manner. In order to reduce the computational burden, this paper uses a hybrid structure which combines a spatial grid of a reduced size at the coarsest level with sub tree below it, down to the finest level. The hidden labels can be estimated by a noniterative inference on sub tree with ICM algorithm defined on the top of the hybrid structure. In order to circumvent the drawback of classical MAP criterion on the hierarchical graph structures, this algorithm combines nonlinear causal Markov model defined on hierarchical graph structures with bayesian SMAP criterion. This algorithm can applied to multi source remote sensed image fusion, and can contribute to the correctness of image pixels labeling and the reduction of the multi source image volume. The contrast experiments on synthetic and satellite images indicated the advantage of this algorithm relative to the classic algorithm.