| 摘要: |
| 近年来深度学习技术在印刷电路板（Printed Circust Boord,PCB）缺陷检测上已获得快速进步，但现有算法针对PCB图像中多尺度高密度微小缺陷目标，如何精准高效地提取特征，提高检测精度及速度依然存在巨大挑战。提出了一种可变形残差卷积与伸缩式特征金字塔的PCB缺陷检测算法。在Faster RCNN的基础上，通过引入可变形残差卷积模块替换原始VGG16网络进行通道关系校准，提高算法对复杂缺陷目标特征的语义获取能力；利用一种伸缩式改进的特征金字塔NAS-FPN网络与原区域建议RPN网络融合，以改善算法对多尺度微小缺陷目标的识别能力；结合IoU Loss、Matrix NMS等tricks组合综合优化网络的检测精度及速度。通过实验，相比原始Faster RCNN，检测精度从90.08%提升到99.41%，检测速率从4.08 frame/s提升到6.47 frame/s。该方法能实现检测精度及速度双高的PCB缺陷检测，具有一定的实际意义。 |
| 关键词: 印刷电路板 缺陷检测 Faster RCNN 可变形卷积 特征金字塔 |
| DOI：10.20079/j.issn.1001-893x.220217002 |
|
| 基金项目:国家自然科学基金资助项目（61305123）；2021年江苏理工学院研究生实践创新项目（XSJCX21_57） |
|
| PCB defect detection based on deformable residual convolution and scalable feature pyramid algorithm |
| SUN Zhichao,WANG Bo,ZHANG Xiaoling |
| (1.School of Electrical Information Engineering,Jiangsu University of Technology,Changzhou 213001,China;2.School of Computer and Software,Dalian Neusoft University of Information,Dalian 116023,China) |
| Abstract: |
| In contemporary years,deep learning techniques have made rapid progress in printed circuit board(PCB) defect detection.Nevertheless,there are still tremendous challenges in how to accurately and efficiently extract features and improve detection accuracy and speed for existing algorithms aiming at multi-scale high-density micro-defect targets in PCB images.The authors propose a PCB defect detection algorithm based on deformable residual and telescopic feature pyramid.On the basis of Faster RCNN,a deformable residual module is introduced to replace the original VG16 network for channel relationship calibration,and the semantic acquisition ability of complex defect target features is improved.An improved feature pyramid NAS-FPN network extension with the original RPN network is adopteed to improve the algorithms recognition ability for multi-scale micro-defect targets.IoU Loss,Matrix NMS and other tricks are used to optimice the detection accuracy and speed of the network.Comparison with the original Faster RCNN,shows the detection accuracy is improved from 90.08% to 99.41%,and the detection rate is improved from 4.08 frame/s to 6.47 frame/s.The method can achieve both high precision and high-speed of PCB defeet detection,so it has practical meaning. |
| Key words: printed circuit board defect detection faster RCNN deformable convolution feature pyramid |
