Application of neural networks in optical inspection and classification of solder joints in surface mount technology

The defect detection on manufactures is extremely important in the optimization of industrial processes; particularly, the visual inspection plays a fundamental role. The visual inspection is often carried out by a human expert. However, new technology features have made this inspection unreliable. For this reason, many researchers have been engaged to develop automatic analysis processes of manufactures and automatic optical inspections in the industrial production of printed circuit boards. Among the defects that could arise in this industrial process, those of the solder joints are very important, because they can lead to an incorrect functioning of the board; moreover, the amount of the solder paste can give some information on the quality of the industrial process. In this paper, a neural network-based automatic optical inspection system for the diagnosis of solder joint defects on printed circuit boards assembled in surface mounting technology is presented. The diagnosis is handled as a pattern recognition problem with a neural network approach. Five types of solder joints have been classified in respect to the amount of solder paste in order to perform the diagnosis with a high recognition rate and a detailed classification able to give information on the quality of the manufacturing process. The images of the boards under test are acquired and then preprocessed to extract the region of interest for the diagnosis. Three types of feature vectors are evaluated from each region of interest, which are the images of the solder joints under test, by exploiting the properties of the wavelet transform and the geometrical characteristics of the preprocessed images. The performances of three different classifiers which are a multilayer perceptron, a linear vector quantization, and a K-nearest neighbor classifier are compared. The n-fold cross-validation has been exploited to select the best architecture for the neural classifiers, while a number of experiments have been devoted to estimating the best value of K in the K-NN. The results have proved that the MLP network fed with the GW-features has the best recognition rate. This approach allows to carry out the diagnosis burden on image processing, feature extraction, and classification algorithms, reducing the cost and the complexity of the acquisition system. In fact, the experimental results suggest that the reason for the high recognition rate in the solder joint classification is due to the proper preprocessing steps followed as well as to the information contents of the features.