印刷电路板(PCB)是保障电子设备产品可靠性的关键因素。因此，对于PCB板的缺陷检测是一项基本和必要的工作。当前PCB缺陷检测方面已经取得了很大进步，但由于PCB板缺陷的多样性、复杂性以及微小性，传统检测方法仍然难以应对。针对PCB板复杂性和微小性问题，文中提出了一种基于深度学习的PCB微小缺陷检测网络命名为UF-Net，该网络通过多层卷积提取不同维度的特征，通过上采样及跳层连接（skip connect）的方式实现多尺度特征融合；然后，利用RPN（Region Proposal Network）网络生成ROI（Region of Interest）；最后，通过ROI-Pooling层提取ROI特征，并经过两个全连接层对ROI区域进行分类和回归，从而实现缺陷检测。文中的方法能够对印制电路板导线缺陷和焊点缺陷进行精确的检测和识别，包括导线的短路、开路、缺口、毛刺以及焊点的孔洞、漏焊、焊盘不全等缺陷。通过对PCB缺陷数据集的定量分析结果表明，该方法具有较好的移植性，在PCB数据集上的精度达到98.6%，满足PCB实际检测需求。
英文摘要:
Printed circuit board (PCB) is the key factor to ensure the reliability of electronic equipment. Therefore, PCB defect detection is a very important job. Although great progress has been made in PCB defect detection, traditional detection methods are still difficult to deal with due to the diversity, complexity, and micro-nature of PCB defects. To solve these problems, a deep learning network for PCB defect detection was proposed and named UF-Net. This network extracts features of different dimensions through multi-layer convolution, and realizes multi-scale feature fusion through up-sampling and skip connect. Then, the Network uses RPN (Region Proposal Network) Network to generate ROI (Region of Interest). Finally, the network classifies and regresses ROI areas through the full connectivity layer to achieve defect detection. Our method can accurately detect and correctly identify the defects of the PCB leads and solder joints, including short circuit, open circuit, mouse bite, the burr of the leads, holes in the solder joints, missing welds, and incomplete welds. The results of the quantitative analysis of PCB defect datasets show that our method has good portability, and the precision of PCB datasets reaches 98.6%, which meets the requirements of practical PCB detection.