基于LVQ神经网络的图像边沿检测研究  

Research on image edge detection using LVQ neural network

作　　者：郭勇[1] 李梦超 谢晓春[1] 乐江源[1] 王兴权[1] 吴元超 GUO Yong;LI Mengchao;XIE Xiaochun;LE Jiangyuan;WANG Xingquan;WU Yuanchao(Physics and Electronic Information,Gannan Normal University School,Ganzhou 341000,China)

机构地区：[1]赣南师范大学物理与电子信息学院,江西赣州341000

出　　处：《光学技术》2021年第4期489-493,共5页Optical Technique

基　　金：江西省教育厅科学技术研究项目(GJJ201437);江西省教育厅科学技术研究项目(GJJ170843)。

摘　　要：分析了传统边沿检测算子的优缺点,引入LVQ神经网络检测图像边沿;阐述了其以传统算子检测结果为教师信号,以灰度图像5×5邻域的中值特征量、方向特征量、Kirsch算子方向特征量为一组特征量作为输入信号训练权值的检测原理与训练过程,给出了特征量的计算公式;以检测脐橙图像边沿为例,设置了不同阈值、不同教师信号类型。结果表明,LVQ神经网络检测边沿,不依赖于教师信号的类型和阈值,在提高边沿连续性和抑制过度曝光点两方面比传统算子检测有显著优势。The advantages and disadvantages of traditional edge detection operators are analyzed;LVQ neural network is introduced to detect image edge;the detection principle and training process of network weight of LVQ is detailed, in which the result of traditional operator detection is taken as teacher signal and the input signal is based on a set of features of 5 × 5 neighborhood, such as median feature, direction feature and Kirsch operator direction feature of gray image, the calculation formula of those features is given;based on different thresholds and teacher signal types, the edge of navel orange image is detected using LVQ neural network;the results show that edge shape of the image is independent on the type and threshold of teacher signal, and as significant advantages over traditional operator detection, edge continuity is improved and overexposure suppressed.

关 键 词：边沿检测 LVQ神经网络 特征量 连续性 过度曝光点 

分 类 号：TH86[机械工程—仪器科学与技术]