摘要
为解决芯片表面缺陷检测的困难,提出新的神经网络模型SE-CPN,引入注意力机制模块,采用自底向上路径聚合的方式检测不同尺度的特征。为便于SE-CPN模型在边缘设备上部署,采用Ghost模块,保证相似的识别性能同时大幅度降低计算量,使存储占用空间降低到16.3 MB。通过各项数据增强手段,在小样本基础上建立了缺陷芯片数据集,训练并评价模型。实验表明,在此数据集上文中模型平均准确率达到98.2%。
To solve the problem of chip surface defect detection, a new neural network SE-CPN model was proposed. Attention mechanism module was introduced to detect different scale features by bottomup path aggregation. To facilitate SE-CPN model deployment on edge devices, Ghost module was used to ensure similar recognition performance while significantly reducing the computational load and reducing the storage space to 16.3 MB. Through various data enhancement methods, a defect chip dataset was established on the basis of small samples, and a training and evaluation model was established. Experiments show that the average model accuracy over this dataset is 98.2%.
作者
夏卓飞
龚家元
周诗薇
代加喜
Xia Zhuofei;Gong Jiayuan;Zhou Shiwei;Dai Jiaxi(School of Automotive Engineers,Hubei University of Automotive Technology,Shiyan 442002,China)
出处
《湖北汽车工业学院学报》
2022年第2期49-54,共6页
Journal of Hubei University Of Automotive Technology
关键词
缺陷检测
数据增强
注意力机制
深度学习
神经网络
defect detection
data enhancement
attention mechanism
deep learning
neural network