为了解决现阶段的印刷电路板(Printed Circuit Board，PCB)缺陷检测方法没有同时关注缺陷的细节信息以及全局信息，跨像素卷积或池化的降采样操作更是造成了PCB表面缺陷全局信息与细节信息的丢失。虽然部分方法使用注意力进行层内信息的关注，但是对普通卷积提取特征后造成的权重偏差问题缺乏关注的问题。本文提出了PCB表面缺陷检测网络(PCB defect detection Network，PCBNet)，该方法通过设计膨胀挤压卷积(Dilation and extrusion Convolution，DeConv)提取PCB表面缺陷全局信息与细节信息，使用空间向通道集中卷积(Spatial to Passage Directed Focused Con-volution，SPD-Conv)进行降采样以减少信息丢失，设计细微特征增强模块(Subtle Feature Enhancement Module，SFEM)调节PCB表面缺陷特征的层内关系以及减少权重偏差的同时增强算法对细微特征的感知能力。在现场采集的PCB表面焊接缺陷数据集以及PCB Defect-Augmented数据集上与多种先进方法进行的对比的实验结果表明，PCBNet不仅在PCB表面焊接缺陷数据集上能够以每秒83帧的速度进行准确识别，还在PCB Defect-Augmented数据集上取得了COCO数据集评价指标mAP0。5的最佳精度。表明本文的方法拥有可部署在嵌入式设备上运行的潜力。
Visual inspection of soldering defects on board surfaces against complex backgrounds
To resolve the current stage of printed circuit board(PCB)defect detection,it is necessary to consider both the detail and global information of the defects simultaneously.The downsampling operation of cross-pixel convolution or pooling results in the loss of both global and detailed information on the sur-face defects of printed circuit boards(PCBs).Although some of the methods above employ attention mechanisms for intra-layer information,the issue of insufficient attention to the weight bias problem result-ing from conventional convolution after feature extraction persists.The PCB defect detection Network Subtle Feature Enhancement Module(SFEM)had been designed to adjust the intra-layer relationship of PCB surface defect features and reduce the weight bias while enhancing the algorithm's ability to perceive the subtle features.The experimental results obtained by comparing the PCB surface soldering defects da-taset and the PCB Defect-Augmented dataset,which were collected in the field using multiple state-of-the-art methods,demonstrate that PCBNet is not only capable of accurately identifying PCB surface soldering defects at a rate of 83 frames per second on the PCB surface soldering defects dataset but also achieves the following results on the PCB Defect-Augmented dataset:the highest accuracy of mAP0.5,which is the evaluation metric of the COCO dataset.This indicates that our method has the potential to be implement-ed on embedded devices.
PCB defect detectiondilation and extrusion convolutionSPD-Convsubtle feature en-hancementobject detection
万方数据
维普
