摘要
针对现有神经网络图像边缘检测算法缺乏足够的参数调节边缘检测的灵敏度,以及检测结果图像边缘过宽的缺陷,提出了一种改进的方法.该方法在神经网络元的n个连接上施加权值,通过各种局部搜索、优化算法和使用指定的样本输入、样本输出等方法,训练加权神经网络确定各权值,在保留了原有算法优点的同时,可以根据不同的样本输入输出图像调节边缘检测的灵敏度,从而提高检测结果质量并避免检测结果边缘过宽的问题.实验结果表明,训练后的加权神经网络方法,有更低的边缘检测错误率,并可检出原有方法漏检的边缘.
In response to the shortcomings of the method of contextual-Hopfield neural network,which are the lack of parameters used to adjust the sensitivity of edge detection and that the edge of image is too wide,an improved method which is called weighted contextual-Hopfield neural network is constructed.This method puts n weights on the n connections of each sub-network.They can be determined by training the network of the new method using local search and optimization methods with the designated sample input and output images.Thus the new network gains the abilities of tuning the sensitivity of edge detection according to different sample images and correcting the too-wide edges detected by contextual-Hopfield neural network.The experimental results show that the trained new network can get lower error rate while processing the images which noise rate is similar to sample input images than the contextual-Hopfield neural network,and it can also detect those methods missed image edge.
出处
《西北师范大学学报（自然科学版）》
CAS
北大核心
2011年第3期40-44,共5页
Journal of Northwest Normal University(Natural Science)
基金
江苏省自然科学基金资助项目(BK2010711)
关键词
边缘检测
人工神经网络
加权参数
参数训练
edge detection
artificial neural network
weighted parameter
parameter training
