基于半监督的电机磁瓦缺陷检测方法  

Defect Detection Method of Motor Magnetic Tile Based on Semi-supervision

作　　者：夏兴华[1] 李欣宇 韩忠华[1] XIA Xinghua;LI Xinyu;HAN Zhonghua(Shenyang Jianzhu University,Shenyang 110168,China)

机构地区：[1]沈阳建筑大学,沈阳110168

出　　处：《移动信息》2024年第9期334-337,共4页Mobile Information

摘　　要：目前大多数的数据集样本都是不含任何标签的,需要耗费大量成本对数据集中的样本进行手动标签.针对这个问题,文中提出了一种基于改进的DCGAN半监督检测模型.首先引入SE注意力机制,通过学习通道间的关系加强有效特征,抑制无效特征,提高网络的学习能力.然后,提出将残差模块与SE注意力机制进行融合应用到生成器和判别器中,减少梯度消失和退化,有效提高生成图片的质量,生成更加逼真的样本数据.最后对损失函数进行优化避免训练过程中的梯度消失或梯度爆炸问题,使模型更容易收敛,且还可以提升生成样本的质量.实验结果显示,改进后的DCGAN半监督检测模型可以减少对全标签数据集的依赖,降低人工成本.At present,most dataset samples do not contain any labels,which requires a significant amount of cost to manually label the samples in the dataset.A semi supervised detection model based on an improved DCGAN is proposed in the article to address this paper.Firstly,the SE attention mechanism is introduced to enhance effective features and suppress ineffective features by learning the relationships between channels,thereby improving the network's learning ability.Then,it is proposed to integrate the residual module with SE attention mechanism and apply it to the generator and discriminator to reduce gradient vanishing and degradation,effectively improve the quality of generated images,and generate more realistic sample data.Finally,the loss function is optimized to avoid gradient vanishing or exploding during the training process,making the model easier to converge and improving the quality of generated samples.The experimental results show that the improved DCGAN semi supervised detection model can reduce dependence on fully labeled datasets and lower labor costs.

关 键 词：缺陷检测 半监督 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]