Gaussian-IoU loss: Better learning for bounding box regression on PCB component detection
Object detection with high accuracy is becoming increasingly important for industrial processes, such as producing printed circuit boards (PCBs). A false or missed detection during production can lead to serious quality issues. Therefore, an efficient detector that can maintain high quality is required for industrial applications. This paper proposes a method for improving detection accuracy while supporting real-time operations using the baseline YOLOv4. A new loss function for box regression called Gaussian intersection of union (GsIoU) is explored, which merges the predicted boxes under different anchors at the same position using a Gaussian function to calculate the box regression loss, improving the accuracy of the final box regression. The proposed PCB Component (PCBC) dataset is a benchmark comprising 18,948 images, 24 categories (containing the same component in different directions), and 508,313 components. Taking YOLOv4 as the experimental baseline on the PCBC dataset, the mean average precision (mAP) using the GsIoU has reached 86.9%, which is improved by 3.3% and 2.2% compared to using the CIoU and GIoU loss functions, respectively. The experiments were conducted on the COCO dataset to verify the generalization of the GsIoU. The detection accuracy of the GsIoU exceeds that GIoU and CIoU by 0.6% on AP50 and 0.6% on AP50:95, achieving 65.6% and 46.0% on the COCO test-val2017, respectively. The detection efficiency of the proposed method is the same as that of the baseline in the testing process and is slightly reduced in the training process owing to the synchronous calculation of the IoU, variance, and Gaussian operations on the network output. The experiments indicate that GsIoU is effective and efficient.
Expert Systems with Applications
Object detection,PCB-components,Industrial inspection,Box regression,Gaussian IoU