A Mask Anomaly Detection Method for Printed Circuit Board Defects

Defect inspection of printed circuit boards (PCBs) is a crucial task in the electronics manufacturing industry, as PCBs contain many components and complex characters. Most of the existing research on this topic relies on the PCB Defect dataset, which only covers six types of defects. However, the defects that occur in industrial production are diverse and variable. It is very challenging to accurately classify all possible defects and obtain a large number of defect training samples. To overcome these difficulties, we propose a novel PCB anomaly detection model called the mask anomaly detection method for printed circuit board defects (MADM-PCB), which is based on the principle of reconstruction error. This model can detect unlimited types of defects on PCBs and only requires positive samples for training, which are more abundant in industry. Moreover, to enhance the detection performance, we introduce Spatial-Shift-MLP (S2-MLP), which can strengthen the connection of adjacent semantics on the channel at low cost. We conduct experiments to demonstrate the effectiveness of our method.