在提取啤酒瓶的缺陷特征后,如何选择合适的多分类支持向量机算法对提高分类准确率和分类速度具有重要的作用.本文通过一对一、一对多、决策有向无环图、二叉树、误差纠错码、一次性求解等多分类支持向量机算法在核函数为线性、多项式、径向基,神经网络的情况下,对多个基准样本进行了分类性能、分类速度、分类准确性的详细比较以及完整的理论分析,最终得出一对一多分类支持向量机在径向基核函数时性能优于其他算法.在啤酒瓶智能检测机器人上的实验,表明这种算法能够满足检测需要.
After extracting defect features, how to choose appropriate multi-class support vector machine classification algorithm is important in the enhancement of classification accuracy rate and classification speed. This paper compared six methods for multi-class SVM classification, i.e. 'One-Against-One', 'One-Against- All', 'Decision Directed Acyclic Graph', 'binary tree', 'Error-eorrecting output coding' and 'considering all data at once' using linear, polynomial, RBF, Sigmoid kernel function. A detailed comparison of several benchmarking samples for classification attribute, classification speed and classification accuracy was conducted through complete theoretical analyses. It has been concluded that 'one Against One' methods using RBF kernel are more superior to the other methods. Experiment results on Beer Bottle inspector have shown that these methods can meet practical requirements.