基于CSTDNet的PCB表面元器件检测算法
PCB surface component detection algorithm based on CSTDNet
郑飞 1储茂祥 1巩荣芬 1刘光虎1
作者信息
- 1. 辽宁科技大学 电子与信息工程学院,辽宁 鞍山 114051
- 折叠
摘要
PCB表面元器件存在分布密集、尺寸小、外观相似等特点,所以检测时容易出现漏检和误检问题.以VFNet为基础,提出了一种名为CSTDNet(Cross-Scale Task Dynamic Network)的PCB表面元器件检测算法.首先,在特征融合网络中添加跨尺度交互特征模块,以增强对小型元器件的特征描述能力;其次,在检测头网络中引入任务对齐学习机制,优化分类和回归任务的空间一致性;另外,在正负样本选择的过程中引入高斯动态软标签分配策略,以更好地补偿小型元器件的正样本数量.实验结果表明,改进后算法的FPS、mAP和mAP_s分别提升了 10.7%、11.8%和 7.6%,有效地提高了密集场景下小型元器件的检测性能.
Abstract
PCB surface components are characterized by dense distribution,small size,and similar appearances,making it challenging to accurately detect and identify issues.Therefore,this paper proposes a Cross-Scale Task Dynamic Network(CSTDNet)PCB surface component detection algorithm based on VFNet.Building upon VFNet,an algorithm named CSTDNet(Cross-Scale Task Dynamic Network)is proposed for the detection of PCB surface components.This algorithm incorporates a cross-scale interactive feature module into the fusion network to enhance the description capability of small components.Furthermore,a task alignment learning mechanism is integrated into the detection head network to optimize spatial consistency between classification and regression tasks.Additionally,a Gaussian dynamic soft label allocation strategy is introduced in the process of positive and negative sample selection to better compensate for the number of positive samples of small components.Experimental results show that the new detection algorithm improves the FPS,mAP,and mAP_s by 10.7%,11.8%and 7.6%,respectively,effectively enhancing the detection performance of small components in dense scenes.
关键词
元器件/VFNet/特征融合/标签分配Key words
components/VFNet/feature fusion/label allocation引用本文复制引用
基金项目
辽宁省自然科学基金(2022-MS-353)
辽宁省教育厅基本科研项目(2020LNZD06)
辽宁省教育厅基本科研项目(LJKMZ20220640)
出版年
2024
NETL
NSTL
万方数据