CNN-IRLS装配孔定位方法研究

Research on CNN-IRLS Assembly Hole Positioning Method

针对零件缺陷、反光或是环境光照不足不均,提出了一种通过卷积神经网络(CNN)一次定位,再二次运用改进迭代重加权最小二乘法(Iterative Reweighted Least Squares,以下简称IRLS)进行筛选和拟合进而进行二次定位的方法。在一次定位时,训练模型的准确率和召回率分别达到98.2%和97.4%,结合二次定位识别率为99.1%,相较于常规形态学筛选和模板匹配在复杂光照下的识别率分别提高了31.9%和15.5%。二次定位时,圆孔的最大定位误差为0.65mm,平均误差0.31mm。对比Hough法和CNN直接定位,最大误差分别减少了33.0%和53.9%,平均误差分别减少了36.7%和50.8%。

Aiming at the defects,reflections or insufficient ambient illumination,a method of secondary positioning is proposed for primary positioning by convolutional neural network(CNN),and then using the improved Iterative Reweighted Least Squares method(hereinafter referred to as IRLS)for screening and fitting.In the first localization,the accuracy and recall rate of the trained model reached 98.2%and 97.4%,respectively,and the recognition rate combined with the secondary positioning was 99.1%,which was 31.9%and 15.5%higher than that of conventional morphological screening and template matching under complex lighting,respectively.In secondary positioning,the maximum positioning error of the round hole is 0.65mm,and the average error is 0.31mm.Compared with the Hough method and CNN direct positioning,the maximum error was reduced by 33.0%and 53.9%,respectively,and the average error was reduced by 36.7%and 50.8%,respectively.