基于Cascade R⁃CNN改进的PCB瑕疵检测方法

Improved PCB defect detection method based on Cascade R⁃CNN

由于印刷电路板(PCB)的集成度、线路复杂性以及其产量的日益增加,针对PCB的瑕疵检测已成为关键的检测任务。传统检测方法存在误检率较大、检测速度较慢、能够检测出的瑕疵类型较少等问题,而常用的基于深度学习的目标检测模型计算量庞大,难以应用到工业场景中算力较弱的边缘设备。为了减少算法参数量和计算量,在Cascade R⁃CNN基础上改进网络结构,提出将ResNeSt作为主干网络,颈部网络采用路径聚合特征金字塔网络(PAFPN),增加对小目标PCB瑕疵的定位精度。将使用多尺度训练方法训练出的高性能改进Cascade R⁃CNN模型作为教师模型,再通过使用重点与全局知识蒸馏算法,将教师模型知识蒸馏在主干轻量化的MobileNetV3学生模型中,在提升模型精度的同时,缓解模型参数量和计算量过大的问题。在PCB Defect数据集上,改进模型相较原Cascade R⁃CNN模型,在IoU=0.5时,mAP提高了17.0%,推理速度提高3.5 FPS,参数量缩小了28.2%,计算量减少了27.3%。

Due to the increasing integration,circuit complexity and output of printed circuit board(PCB),defect detection of PCB has become the most important detection task in the entire electronic industry.The traditional detection methods have the problems of large false detection rate,slow detection speed,and few types of defects that can be detected.The commonly used object detection model based on deep learning has a large amount of calculation and is difficult to apply to edge devices with weak computing power in industrial scenes.In order to reduce the amount of algorithm parameters and calculations,the network structure is improved on the basis of Cascade R⁃CNN,and ResNeSt is proposed as the backbone network.The path aggregation feature pyramid network(PAFPN)is used in neck network to increase the positioning accuracy of small target PCB defects.The high⁃performance improved Cascade R⁃CNN model trained by multi⁃scale training method is used as the teacher model.The key and global knowledge distillation algorithm is used to distil the teacher model knowledge into the backbone lightweight MobileNetV3 student model,which can alleviate the problem of excessive large model parameters and computation while improving model accuracy.On the PCB Defect dataset,in comparison with the original Cascade R⁃CNN model,when IoU=0.5,the improved model can increase mAP and inference speed by 17.0%and 3.5 FPS,respectively,and reduce parameter and computation by 28.2%and 27.3%,respectively.