李志诚,曾志强.基于改进YOLOv3的卷纸包装缺陷实时检测算法[J].中国造纸学报,2022,(2):87-93 本文二维码信息
二维码(扫一下试试看!)
基于改进YOLOv3的卷纸包装缺陷实时检测算法
Real-time Defect Detection Algorithm for Roll Paper Packaging based on Improved YOLOv3
投稿时间:2021-07-15  
DOI:10.11981/j.issn.1000-6842.2022.02.87
中文关键词:  卷纸包装  缺陷检测  卷积神经网络  自注意力  多尺度特征
Key Words:roll paper packaging  defect detection  convolutional neural networks  self-attention  multi-scale features
基金项目:广东省基础与应用基础研究基金(2020A1515011468);广东省普通高校特色创新类项目(2019KTSCX189)。
作者单位邮编
李志诚 五邑大学智能制造学部广东江门529000 529000
曾志强 五邑大学智能制造学部广东江门529000 529000
摘要点击次数: 562
全文下载次数: 586
中文摘要:
      为了解决卷纸包装过程中的缺陷检测问题,提高卷纸包装缺陷的检测速度和准确率,提出了一种改进的YOLOv3算法(iYOLOv3算法)。该算法通过将卷积神经网络和多头自注意力相结合,可更加充分地提取图片的局部和全局特征;并将不同尺度的特征图进行多尺度特征融合,同时改善了YOLOv3算法的解码公式,iYOLOv3算法的AP@50∶5∶95比YOLOv3算法提高了5.8个百分点,检测速度达到了80帧/s,是YOLOv3算法的2倍以上,表明了其在实际应用的可行性。
Abstract:
      In order to solve the problem of defect detection in the process of roll paper packaging and improve the detection speed and accuracy of roll paper packaging defects, an improved YOLOv3 (iYOLOv3) algorithm was proposed. By combining convolutional neural network and multi-head self-attention, the iYOLOv3 algorithm could extract partial and global features of the image more adequately and it could further fuse the feature maps with different scales in multi-scale so that the decoding formular of YOLOv3 algorithm was improved: the AP@50∶5∶95 of iYOLOv3 algorithm was 5.8 percentage points higher than that of YOLOv3 algorithm, and its detection speed reached 80 frames/s, 2-folds more than that of the YOLOv3 algorithm, indicating its feasibility in practical application.
查看全文  查看/发表评论  下载PDF阅读器  HTML

分享按钮