摘要
该文提出一种基于卷积神经网络的模板重建方法,采用残差学习方式逐级精细化得到重建结果,通过产品图像与模板的比对完成对工业品的外观质量检测。在模板重建过程中,结合自注意力机制的关联度检索与编码融合方式,在保持细节还原效果的同时大幅减少了计算量;并提出域适应对抗学习方法,避免重建过程对缺陷信息的还原,显著控制了检测漏报率。实验结果表明了该方法的有效性与较强适应能力。
This paper proposes a template reconstruction method based on convolutional neural network,which uses residual learning method to refine the reconstruction results step by step,and completes the appearance quality inspection of industrial products by comparing product images with templates.In the process of template reconstruction,adopting the self-attention mechanism via relevance retrieval and codes fusion method greatly cuts down the calculation while maintaining the advantage of detail restoration.A domainadaptive adversarial learning strategy is further proposed to avoid the restoration of defect information in the reconstruction procedure,and significantly control the failure rate of defect detection.The experimental results prove the effectiveness and adaptability of this approach.
作者
贾可
赵锞
曾欣科
贾力
李孝杰
JIA Ke;ZHAO Ke;ZENG Xinke;JIA Li;LI Xiaojie(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)
出处
《现代信息科技》
2020年第18期1-6,共6页
Modern Information Technology
基金
四川省科技计划资助项目(2019YFG0189)。
关键词
卷积神经网络
模板重建
缺陷检测
自注意力
域适应对抗学习
convolutional neural network
template reconstruction
defect detection
self-attention
domain-adaptive adversarial learning
