采用多分类器集成方法的带钢表面缺陷图像识别

现有带钢表面缺陷在线识别系统中单个分类器对部分缺陷识别率不高,并且对训练样本依赖性较大;针对这一问题,提出了一种基于并行多分类器集成技术的带钢缺陷图像识别方法.该方法选择LVQ神经网络、RBF神经网络和支持向量机作为基分类器,应用加权投票法对基分类器进行集成,从而实现基分类器能力互补.实验表明,采用多分类器集成的带钢表面缺陷图像识别方法可以更准确地对带钢常出现的边缘锯齿、焊缝、夹杂、抬头纹等缺陷进行识别,能够得到相当或优于任何单个分类器的分类精度,总体识别率达到96%以上.

Surface Defect Recognition for Steel Strips by Combining Multiple Classifiers

To solve the problems that a single classifier recognizes the surface defects of steel strips ineffectively and over-depends on training samples,a new method with combination of multiple classifiers was proposed.The LVQ and RBF neural networks,and the support vector machine were used as the basic classifiers.The weighted voting algorithm was applied to integrating these basic classifiers,thus the complementary of the recognition system was realized.The experiments showed that the common surface defects of steel strips such as zigzag edges,welding seams,inclusions and wrinkles can be more effectively recognized by the combined multiple classifiers.The classification accuracy is better than that of a single classifier,with the overall recognition rate above 96%.