【摘要】 在PCB贴片安装的机器视觉检测中,引脚焊接缺陷检测由于图像数据量大、变化复杂、样本和检测错误代价不平衡及检测的实时性要求,检测难度很大。从图像中选择和提取多种特征,分别用于各个简单的ANN检测分类器,再将多个ANN进行线性集成来得到最后的检测结果可以提高检测精度和速度。各ANN学习时,样本初始权重考虑样本的不平衡性,学习中再用AdaBoost算法来调整样本权重和集成系数;用遗传算法来学习确定ANN,用代权重的检测正确率和最小正确分隔边缘作为适应值函数,两类边缘与代价成正比。实验结果表明:该方案准确率高、泛化性好、速度快。更多还原

【Abstract】 In automatic machine vision inspection for PCB SMT assembly,components solder joints inspection was a meaningful but hard work suffered from various complicated mass image data and unbalanced samples.This paper proposed to solve it by ANN ensembles each of which input by a different feature extracted from the image.Every simple ANN was trained by genetic algorithm whose fitness function was exactness with loss-weighted minimal margin.AdaBoost algorithm provided ANN linear combination coefficients and training samples’ weights were adjusted according to other ANN’s training results.Experimental results showed that this approach had high accuracy,well generalization performance and low calculation cost.更多还原