谢钰宁  叶廷东  陈燕  升谢峰

（广东轻工职业技术大学人工智能学院，广东 广州 510300）

摘要：针对印制电路板（PCB）小尺度缺陷在整图推理中响应较弱、易被固定工作阈值过滤，以及局部复检结果直接融合易引入误检的问题，提出一种参考模板引导的PCB缺陷选择性局部复检方法。该方法分别在待检测PCB图像、模板差异增强图上进行整图检测，并将两路整图检测输出逐类融合后按工作阈值进行筛选，得到基础整图融合结果；同时，采用候选阈值对两路整图检测输出进行筛选、候选融合，以保留弱响应候选区域；每幅图像最多选择2个候选区域进行局部复检；最后，结合基础整图融合结果、触发候选的类别、空间重叠关系、专家输出置信度，对专家输出进行筛选与融合。在DeepPCB测试集上的实验结果表明，本文方法的召回率、F1分别为98.21%、97.27%，较基础检测器方法分别提高了4.65和2.42个百分点，漏检数减少了70个，表明该方法能够在较低的额外推理成本下补充弱响应漏检并抑制新增误检。

关键词：印制电路板；缺陷检测；参考模板；局部复检；误检抑制

中图分类号：TP391.41           文献标志码：A          文章编号：1674-2605(2026)04-0006-08

DOI：10.12475/aie.20260406                            开放获取

Reference Template-guided Selective Local Re-inspection Method for PCB Defects

XIE Yuning  YE Tingdong  CHEN Yansheng  XIE Feng

(Guangdong Industry Polytechnic University, College of Artificial Intelligence, Guangzhou 510300, China)

Abstract: To address the issues that small-scale defects on printed circuit boards (PCBs) exhibit weak responses in whole-image inference and are easily filtered out by fixed working thresholds, as well as the problem that direct fusion of local re-inspection results tends to introduce false alarms, this paper proposes a reference template-guided selective local re-inspection method for PCB defects. The method performs whole-image detection on both the test PCB image and the template-difference enhanced image, then fuses the two whole-image detection outputs in a class-wise manner and filters them with a working threshold to obtain a base whole-image fusion result. Meanwhile, a candidate threshold is used to filter and fuse the two whole-image detection outputs to retain weak-response candidate regions. For each image, at most two candidate regions are selected for local re-inspection. Finally, the expert outputs are screened and fused based on the base whole-image fusion result, the triggered category, spatial overlap relationships, and the confidence of the expert outputs. Experimental results on the DeepPCB test set show that the proposed method achieves a recall of 98.21% and an F1-score of 97.27%, which are 4.65 and 2.42 percentage points higher than those of the baseline detector, respectively, and reduces the number of missed detections by 70, indicating that the method can effectively supplement weak-response missed detections and suppress newly introduced false alarms at low additional inference cost.

Keywords: printed circuit board (PCB); defect detection; reference template; local re-inspection; false detection suppression