Journal of Computer Applications ›› 2016, Vol. 36 ›› Issue (10): 2832-2836.DOI: 10.11772/j.issn.1001-9081.2016.10.2832

Previous Articles     Next Articles

Supervised active contour image segmentation by kernel self-organizing map

FAN Haiju1,2, LIU Guoqi2   

  1. 1. College of Navigation and Aerospace Target Engineering, Information Engineering University, Zhengzhou Henan 450002, China;
    2. College of Computer and Information Engineering, Henan Normal University, Xinxiang Henan 453007, China
  • Received:2016-03-25 Revised:2016-06-15 Published:2016-10-10
  • Supported by:
    BackgroundThis work is partially supported by the National Natural Science Foundation of China (U1404603), the Key Scientific and Technological Research Project of the Provincial Education Department of Henan Province (13A520522).

基于核自组织映射的有监督主动轮廓图像分割

范海菊1,2, 刘国奇2   

  1. 1. 信息工程大学 导航与空天目标工程学院, 郑州 450002;
    2. 河南师范大学 计算机与信息工程学院, 河南 新乡 453007
  • 通讯作者: 范海菊,E-mail:13781908706@163.com
  • 作者简介:范海菊(1979—),女,河南新乡人,副教授,博士研究生,主要研究方向:图像分割、计算机视觉;刘国奇(1984—),男,河南新乡人,副教授,博士,主要研究方向:图像分割、计算机视觉。
  • 基金资助:
    国家自然科学基金资助项目(U1404603);河南省教育厅科学技术重点研究项目(13A520522)。

Abstract: The objects with inhomogeneous intensity or multi-gray intensity by using active contour, a supervised active contour algorithm named KSOAC was proposed based on Kernel Self-Organizing Map (KSOM). Firstly, prior examples extracted from foreground and background were input into KSOM for training respectively, and two topographic maps of input patterns were obtained to characterize their distribution and get the synaptic weight vector. Secondly, the average training error of unit pixel of two maps were computed and added to energy function to modify the contour evolution; meanwhile, the controlling parameter of energy item was obtained by the area ratio of foreground and background. Finally, supervised active contour energy function and iterative equation integrated with synaptic weight vectors were deduced, and simulation experiments were conducted on multiple images using Matlab 7.11.0. Experimental results and simulation data show that the map obtained by KSOM is closer to prior example distribution in comparison with Self-Organizing Map (SOM) active contour (SOAC), and the fitting error is smaller. The Precision, Recall and F-measure metrics of KSOAC are higher than 0.9, and the segmentation results are closer to the target; while the time consumption of KSOAC is similar to that of SOAC. Theoretical analysis and simulation results show that KSOAC can improve segmentation effectiveness and reduce target leak in segmenting images with inhomogeneous intensity and objects characterized by many different intensities, especially in segmenting unknown probability distribution images.

Key words: Kernel Self-Organizing Map (KSOM), active contour, energy function, image segmentation, intensity inhomogeneity

摘要: 针对灰度不均匀目标和多灰度强度目标利用主动轮廓难以精确分割的问题,提出了一种基于核自组织映射(KSOM)的有监督主动轮廓算法KSOAC。首先对背景区域和前景区域的先验样本分别利用KSOM进行训练,得到其各自的拓扑映射结构来表征其分布,从而获得突触权值向量;其次提出计算两个网络结构单位像素的平均训练误差,把该误差加入能量函数修正曲线进化过程,并利用前景和背景的面积比得出能量项的控制参数;最后推导出了利用神经元权值向量的有监督主动轮廓能量函数和迭代方程,并采用Matlab 7.11.0对多幅图像进行了仿真验证。仿真结果和数据表明,与自组织映射(SOM)主动轮廓(SOAC)相比,KSOM得到的映射更接近于先验样本的分布,误差更小;KSOAC的准确率、查全率和F参数均大于0.9,分割结果更接近目标本身;在时间消耗方面与SOAC相差不大。实验结果表明,KSOAC能够提高概率分布未知图像、非均匀图像和多灰度强度目标分割效果,减少目标泄露。

关键词: 核自组织映射, 主动轮廓, 能量函数, 图像分割, 灰度不均匀

CLC Number: