基于核自组织映射与图论的图像分割方法

Image segmentation based on kernel self-organizing mapping and graph theory

为了对以特征聚类为基础的图像分割方法进行目标优化并提高分割性能,提出了一种核自组织映射与EGB(efficient graph-based)算法相结合的自适应分割方法。将依据信息理论推导出的核自组织映射应用于图像分割,使得图像经映射聚类后,同一分类内像素的相似度最高且信息熵最大,不同分类间的互信息最小,从而得到最符合图像分割目标的聚类效果。将聚类得到的区域进一步用改进的EGB算法自适应地进行合并,既充分结合了像素的空间特性,又能克服EGB算法的不足,可获得非常准确的分割结果。在综合分析多种图像分割评价方法的基础上,选取了一些量化指标对分割结果进行客观评价。实验及分析结果表明,本文的分割方法准确可靠,其图像分割结果的量化评价指标明显优于EDISON方法。

In order to optimize the objective functions of the image segmentation methods based on feature clustering and improve their performance, an adaptive segmentation method combining kernel self-organizing mapping and efficient graph-based （EGB） algorithm is proposed. In our approach, the kernel-based topographic mapping derived from information theory is applied to optimize the objectives of segmentation. When using this topographic mapping to cluster the image, the similarity of pixels within the same cluster and the differential entropy of each cluster are maximized,and the mutual information between different clusters is minimized. This leads to the best image clustering. According to the clustering results, the image is further merged by the improved EGB algorithm adaptively. This helps to make full use of the image＇s spatial information and avoid the shortcoming of the algorithm. On the basis of comprehensive analysis of a variety of segmentation evaluation methods, we select some quantitative evaluations to judge the performance of our method objectively. Experimental and analytical results show that our method is accurate and reliable,and its quantitative evaluation results are superior to those of the EDISON method.