基于ResNeSt的烟支外观缺陷分类

Classification of cigarette appearance defects based on ResNeSt

自动烟支外观缺陷分类是卷烟厂高速流水线上产品质量检测需要解决的问题,是提高卷烟品质和竞争力的有效措施.基于烟草企业的实际需求,提出了一种基于ResNeSt模型的烟支外观缺陷分类方法.首先,针对烟支外观缺陷样本数量不足的问题,采用了迁移学习的方法;其次,针对烟支图像的特征,采用多尺度测试,输入不同尺度大小的图片进行训练;最后,为了更好地提取缺陷特征,提高分类准确率,用h-swish替换ReLU激活函数.实验结果显示,准确率达到了92.04%,提出的方法比另外10种主流网络在分类准确率上更高.

In the product quality inspection of cigarette factories,automatic classification of cigarette appearance defect is a problem that needs to be solved on high-speed assembly lines.It is an effective measure to improve the quality and competitiveness of cigarettes.Deep learning is used to classify cigarette appearance defects,and a classification method of cigarette appearance defect based on ResNeSt is proposed in this paper.Firstly,the transfer learning is adopted to solve insufficient samples of cigarette appearance defects.Secondly,according to the characteristics of cigarette images,multi-scale testing is used,and different scales images are inputted to train.Finally,h-swish is used to replace the ReLU activation function in order to better extract the defect features and improve the classification accuracy.The experimental results show that the proposed method has a higher classification accuracy than the other 10 mainstream networks.