针对PCB产品视觉检测中图像缺陷细微、形状复杂、特征难于提取、易受噪声影响的问题,提出基于小波变换和光滑支持向量机集成SSVME(Smooth Support Vector Machine Ensemble)的多分类方法,有效解决了细微、复杂缺陷难以识别分类的问题。实验表明,该方法六类缺陷混合识别率达到95.26%,高于BP神经网络的最优识别率90.35%和基于区域方法的80.67%,而且训练和分类时间短。从理论和实验中验证了该方法的有效性,是PCB产品视觉检测领域中缺陷识别分类的新方法,具有重要的应用价值。
Regression analysis is often formulated as an optimization problem with squared loss functions. Facing the challenge of the selection of the proper function class with polynomial smooth techniques applied to support vector regression models, this study takes cubic spline interpolation to generate a new polynomial smooth function ｜×｜ε^ 2, in g-insensitive support vector regression. Theoretical analysis shows that Sε^2 -function is better than pε^2 -function in properties, and the approximation accuracy of the proposed smoothing function is two order higher than that of classical pε^2 -function. The experimental data shows the efficiency of the new approach.
