基于SIFT特征与多层BP神经网络的钢板缺陷检测算法  

Steel Plate Defection Detection Algorithm Based on SIFT Feature and Multi-Layer BP Neural Network

作　　者：朱晓珺[1] 韩林 邹香玲 

机构地区：[1]河南广播电视大学信息工程学院,郑州450000

出　　处：《组合机床与自动化加工技术》2017年第10期54-56,61,共4页Modular Machine Tool & Automatic Manufacturing Technique

基　　金："核高基"国家科技重大专项(2009ZX01036);河南省科技攻关项目(152102310377)

摘　　要：为了解决当前钢板表面缺陷在对比度弱、边缘复杂和光照不均干扰下易导致检测能力较低的问题,文章提出了基于SIFT特征与多层BP神经网络的钢板缺陷检测算法。首先,引入高斯差分和Hessian矩阵,对钢板图像进行空间尺度函数计算,统计SIFT深度向量特征,完成缺陷特征的检测与收集。然后,基于神经网络原始模型,计算其第五层输出结果,优化缺陷检测结果,并最小化输出层与期望值的差异平方,滤除伪SIFT特征的干扰,建立多层BP神经网络拓扑分析算子,准确识别钢板缺陷。最后,基于软件工程,设计检测系统软件,对文中算法的缺陷检测精度进行测试。实验测试结果显示:与当前主流钢板缺陷检测技术相比,文中算法拥有更高的准确性与鲁棒性。In order to solve the current steel plate surface defect contrast w eak edge,complex and uneven illumination,lead to the problem of insufficient recognition detection algorithm,the steel plate defection detection algorithm based on SIFT feature and multi-layer BP neural netw ork w as proposed. Firstly,the Gaussian difference and Hessian matrix w ere introduced to calculate the spatial scale function for statistical SIFT depth vector features to complete defect detection and collection. Then,the output of fifth layers is calculated for filtering the interference of pseudo SIFT feature,and multilayer BP neural netw ork topology analysis operator w as established to achieve the purpose of accurate identification plate defects. Finally,detection system softw are w as designed based on the softw are engineering to test the defect detection accuracy of this algorithm. The experimental results show that the algorithm is more accurate and robust compared w ith the current steel plate defection detection technology.

关 键 词：钢板缺陷检测 SIFT特征 BP神经网络 空间尺度 深度向量 

分 类 号：TH162[机械工程—机械制造及自动化] TG506[金属学及工艺—金属切削加工及机床]