摘要
文章针对机器视觉表面缺陷检测中不同类缺陷样本数量少和不均衡的情况,提出了用1-分类分别对单独类缺陷进行真/假分类判断的分类方法,首先对每类训练样本计算具有尺度和旋转不变的不变矩特征,再使用基于1-分类的支持向量机和RBF核函数对每一类缺陷样本生成一个超球面,然后通过二重网格搜索的方法对核函数的参数寻优,最后对实际采集的缺陷图像自动寻找缺陷位置并进行分类。实验表明,1-分类支持向量机进行缺陷分类能克服分类样本不均衡的限制,具有分类准确率高及易实现在线检测等优点。
A method based on 1-class classification is proposed for single class defects true/false judg- ments when training samples are less or sample numbers are different for different classes. First, the moment invariant feature which has scale and rotation invariance for training samples of every class is calculated. Then, the 1-class support vector machine(SVM) based on RBF kernel function is used to generate a hypersphere for every defects class. After that, a dual grids searching method is proposed to find parameters optimization of kernel function. Finally, the defects on the actual captured image are located and classified automatically. The results show that 1-class SVM method for defects classi- fication can take advantages of unbalanced samples distribution between different classes, and it has high precision and is easy for on-line inspection.
出处
《合肥工业大学学报(自然科学版)》
CAS
CSCD
北大核心
2012年第10期1311-1315,共5页
Journal of Hefei University of Technology:Natural Science
关键词
缺陷检测
不均衡样本
1-分类SVM
RBF核函数
二重网格
defect inspection
unbalanced sample
1-class support vector machine(SVM)
RBF kernel function
dual grid