An efficient SMD-PCBA detection based on YOLOv7 network model

Modern Printed Circuit Board Assembly (PCBA) manufacturing processes require more accurate and robust defect inspection methods. Despite the potential of deep learning algorithms in PCBA defect detection, their ability to handle environmental factors and multi-modal data is still limited. To overcome this, we propose an improved YOLOv7-based network model that enhances the detection performance of densely distributed multi-size Surface Mount Devices (SMD) in multi-modal PCBA. Specifically, the proposed model enhances feature representation by designing a detection head based on Coordinate Attention, and incorporates feedback connections in the feature fusion stage to improve feature recognition through low-level propagation. Additionally, we propose the SEIoU loss function to calculate position loss between the prediction box and the ground truth, resulting in superior regression accuracy of the anchor box and improved detection accuracy. We validate the effectiveness and improvement of our proposed method through ablation experiments and algorithm comparison. Our proposed model outperforms the baseline YOLOv7 model with a 2.1% increase in [email protected] for the multi-modal PCBA dataset and a 4.5% increase in [email protected]:0.95 for the VOC 2012 dataset, and a 1.0% increase in [email protected]:0.95 for the COCO 2017 dataset. Our study's results suggest that our proposed model is a promising alternative to existing methods for detecting PCBA defects, as it accurately detects multiple tiny components amidst complex backgrounds, effectively identifies diverse types of defects, and remains lightweight.