基于SOFM神经网络的图像融合二值化方法  

Image fusion binarization method based on SOFM neural network

作　　者：潘梅森[1] 荣秋生[1] 

机构地区：[1]湖南文理学院计算机科学与技术系,湖南常德415000

出　　处：《光学精密工程》2007年第3期401-406,共6页Optics and Precision Engineering

基　　金：湖南省教育厅资助科研项目(No.05C720)

摘　　要：提出了一种基于自组织特征映射(SOFM)神经网络的图像融合二值化方法。介绍了SOFM神经网络的特点及学习算法,根据SOFM的聚类确定图像第一阈值作为循环迭代的初始值,对整幅图像进行循环迭代得到第二阈值,使用第二阈值对原始图像进行二值化,得到第一幅待融合图像;通过改进的Bernsen方法对原始图像进行二值化,得到第二幅待融合图像;最后根据图像灰度值选小的原则作为图像融合方法,得到最终的二值化图像。该方法既能有效地消除伪影,又能较好地分离字符和文字。模拟实验结果表明,该方法的二值化效果明显优于Bernsen方法和Ostu方法,且具有良好的适应性。An image fusion binarization method based on Self-organization Feature Map（SOFM） neural network is presented. The characteristic and the learning algorithm of SOFM neural network are introduced. The first threshold of the image is derived from the clustering characteristic of SOFM, and it is treated as a initialization of the circulated iteration method to obtain the second threshold of the image. The binarization for the original image is carried on using the second threshold to obtain the first fusing image. Then, the binarization for the original image is performed more again using the improved Bernsen method to obtain the second fusing image. Finally, two fusing images are merged based on the minimum principle of image gray values to get the final binarization image. This method can effectively eliminate the ghost and can also separate the characters very well. Experimental results show that the method is effective, and its binarization effect surpasses Bernsen method and Ostu method obviously, moreover it has good compatibility.

关 键 词：图像融合 二值化 闽值 SOFM神经网络 像素 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]