SMT Assembly Inspection Using Dual-Stream Convolutional Networks and Two Solder Regions

The automated optical inspection of a surface mount technology line inspects a printed circuit board for quality assurance, and subsequently classifies the chip assembly defects. However, it is difficult to improve the accuracy of previous defect classification methods using full chip component images with single-stream convolutional neural networks due to interference elements such as silk lines included in a printed circuit board image. This paper proposes a late-merge dual-stream convolutional neural network to increase the classification accuracy. Two solder regions are extracted from a printed circuit board image and are input to a convolutional neural network with a merge stage. A new convolutional neural network structure is then proposed that is able to classify for defects. Since defect features are concentrated in solder regions, the classification accuracy is increased. In addition, the network weight is reduced due to a reduction of the input data. Experimental results for the proposed method show a 5.3% higher performance in F1-score than a single-stream convolutional neural network based on full chip component images.