ANN Ensembles Based Machine Vision Inspection for Solder Joints

In automatic machine vision inspection for PCB SMT assembly, solder joints inspection is a meaningful but difficult work suffered from variety of complicated images and imbalance of samples. This paper proposed to use ANN ensembles for inspecting classifier, and every simple ANN for ensembles was input with a different feature selected from the image and was trained by genetic algorithm whose fitness function was weighted correctness combined with the minimal margin. ANN linear combination coefficients and training samples´ weights were adjusted by AdaBoost algorithm, which corrected a former trained ANN´s errors by training other ANNs. Experimental results showed that this approach had got high accuracy, well generalization performance and low calculation cost.