SD2Net: surface-mounted device detection network with convolution-free attention mechanism for printed circuit board integrity assurance

Abstract

The accuracy of detecting and classifying surface-mounted devices (SMDs) on printed circuit boards (PCBs) is influenced by the complexity of the background and the variations in scale among the SMDs. Several object detection networks, such as the you only look once (YOLO) series-based networks and two-stage models like Mask R-CNN, have been developed to handle the detection of surface-mounted devices (SMDs) at various scales. However, the majority of these networks exhibit a complex structure with an abundance of parameters, which presents difficulties in deploying SMD detection networks in real-world application scenarios. To overcome this challenge, we introduce SD2Net, a convolution-free attention-based network specifically designed for accurate SMD detection. The front-end of SD2Net employs the proposed pyramid vision transformer (PVT) block-based backbone to extract SMD features in five stages, incorporating an attention mechanism. Subsequently, the extracted multi-scale features are fed into the feature pyramid network (FPN) for efficient feature fusion. In the back-end of SD2Net, the final stage integrates the VarifocalNet head module (VFH). By utilizing the Varifocal loss and generalized intersection over union (GIoU), it performs precise regression and computation of the localization and class mapping vector for SMDs. We conduct extensive experiments using two well-established publicly available datasets: FICS-PCB and PCB-WACV. The experimental results indicate the superior performance of our model in comparison to state-of-the-art methods.