Printing circuit board (PCB) defect inspection precisely and efficiently is an essential and challenging issue. Therefore, based on several improvements upon YOLOv5-nano, a novel lightweight detector named TD-YOLO is proposed to inspect tiny defects in PCBs. First, the lightweight ShuffleNet block is implemented into the backbone to effectively reduce the model weight. Second, novel anchors are designed using modified k-means clustering to accelerate the model convergence and yield superior detection precision. Then, data augmentation strategy is recomposed by rejecting mosaic augmentation to suppress the emergence of extremely tiny targets. Finally, a mighty feature pyramid network namely MPANet, is newly proposed to boost the feature fusion capability of the model. The experiment results denote TD-YOLO achieves the highest 99.5% mean average precision on our dataset, outperforming other state of the arts. Specially, the detection metrics for the smallest two defects, such as spur and mouse bite, are increased by 2.1% and 1.2%, respectively, compared with YOLOv5-nano. Besides, TD-YOLO has only 1.33 million parameters, decreased by 25% than the baseline. Using a mediocre processor, the detection speed is boosted by 20%, reaching 37 frames per second for the input size of 2240$\times$2240 pixels.

中文翻译：

---

TD-YOLO：一种用于高分辨率 PCB 微小缺陷的轻量级检测算法

精确、高效的印刷电路板（PCB）缺陷检测是一个重要且具有挑战性的问题。因此，基于YOLOv5-nano的多项改进，提出了一种名为TD-YOLO的新型轻量级检测器来检查PCB中的微小缺陷。首先，将轻量级的ShuffleNet块实现到主干中，以有效降低模型权重。其次，使用改进的 k 均值聚类设计新颖的锚点，以加速模型收敛并产生卓越的检测精度。然后，通过拒绝马赛克增强来重构数据增强策略，以抑制极小目标的出现。最后，新提出了一个强大的特征金字塔网络，即MPANet，以提高模型的特征融合能力。实验结果表明 TD-YOLO 在我们的数据集上达到了最高的 99.5% 平均精度，优于其他现有技术。特别是，与YOLOv5-nano相比，最小的两个缺陷（例如毛刺和老鼠咬伤）的检测指标分别提高了2.1%和1.2%。此外，TD-YOLO只有133万个参数，比基线减少了25%。使用平庸的处理器，检测速度提升了20%，对于2240的输入大小达到每秒37帧$\times$2240 像素。