摘要
针对印刷电路板(printed circuit board,PCB)计算机断层成像(computerized tomography,CT)图像通常存在噪声大、过孔和焊盘的数量多以及焊盘与背景的对比度低和焊盘形状变化较多等因素导致过孔和焊盘的检测比较困难的问题,基于卷积神经网络模型网中网,提出将池化核作为参数进行学习以提高网络的数据表达能力,在基准数据集上进行验证后结合选择性搜索算法应用于PCB CT图像中的过孔和焊盘检测。实验结果表明,参数池化能够帮助提高网络对数据的表达能力,改进后的网络能够有效检测出PCB CT图像中的过孔和焊盘,基本达到实际应用需求。
Aiming at the problems of large noise, large amount of vias and pads, low contrast between pads and backgrounds, and large variation of pads＇ shape and scale in PCB CT images, which make it difficult to detect vias and pads. Based on the convolutional neural network model network in network, this paper proposed to parameterize the pooling kernels to improve the network＇ s data-expressing ability, which was applied to via and pad detection combined with selective search algorithm after evaluated on benchmark datasets. The experiments show that, parameterized pooling can help improve the network＇ s data-expressing ability, with which the convolutional neural network can effectively detect vias and pads in PCB CT images and basically meet the practical need.
出处
《计算机应用研究》
CSCD
北大核心
2018年第2期637-640,共4页
Application Research of Computers
基金
国家自然科学基金资助项目(61372172)
