河南农业科学 ›› 2022, Vol. 51 ›› Issue (11): 145-154.DOI: 10.15933/j.cnki.1004-3268.2022.11.017

• 农业信息与工程·农产品加工 • 上一篇    下一篇

基于高效通道注意力机制与多尺度特征融合的烟丝图像识别方法研究

刘江鹏1,牛群峰1,靳毅2,陈霞2,王莉1,袁强1   

  1. (1.河南工业大学 电气工程学院,河南 郑州 450000;2.河南中烟工业有限责任公司安阳卷烟厂,河南 安阳 455006)
  • 收稿日期:2022-08-01 出版日期:2022-11-15 发布日期:2023-01-10
  • 通讯作者: 牛群峰(1974-),男,河南郑州人,副教授,博士,主要从事智慧农业及农业信息化技术研究。E-mail:niuqunfeng@163.com
  • 作者简介:刘江鹏(1998-),男,重庆黔江人,在读硕士研究生,研究方向:机器学习及计算机视觉等技术。E-mail:ljpcontrol@163.com
  • 基金资助:
    河南省科技攻关计划项目(201300210100,222102110160);河南中烟科技项目(A202047,A202049)

Research on Tobacco Shred Image Recognition Method Based on Efficient Channel Attention Mechanism and Multi‑Scale Feature Fusion

LIU Jiangpeng1,NIU Qunfeng1,JIN Yi2,CHEN Xia2,WANG Li1,YUAN Qiang1   

  1. (1.College of Electrical Engineering,Henan University of Technology,Zhengzhou 450000,China;2.Anyang Cigarette Factory,China Tobacco Henan Industry Co.,Ltd.,Anyang 455006,China)
  • Received:2022-08-01 Published:2022-11-15 Online:2023-01-10

摘要: 针对现有方法在识别烟丝类型中泛化能力差、准确率低的问题,提出了一种基于高效通道注意力机制与多尺度特征融合的烟丝类型识别方法。对采集的梗丝、膨胀叶丝、叶丝和再造烟丝4类烟丝图像进行降噪处理,处理后的图像经K-means聚类得到图像的前景和后景并完成分割,提高输入图像的抗环境干扰能力和特征提取能力。在Inception-ResNet-V2网络中引入高效通道注意力机制,加强模型提取特征的能力;同时,将改进后的模块输出的特征图进行多尺度融合,增加特征代表性,降低过拟合风险。最后,在比较收敛性和准确性时,用PReLU和AdaBound代替了ReLU激活函数和Adam优化器。结果表明,提出的算法具有较好的泛化能力,能实现4类烟丝高效识别,最终识别精度为97.23%,单幅图像的检测时间为0.107 s。

关键词: 烟丝, k-means算法, Inception网络, 高效通道注意力机制, 多尺度特征融合

Abstract: To address the problems of poor generalization ability and low accuracy of existing methods in identifying tobacco shred,a method for identifying tobacco shred types based on efficient channel attention mechanism and multi‑scale feature fusion is proposed.Noise reduction is performed on the four types of collected tobacco images,cut stem,expanded tobacco silk,tobacco silk and reconstituted tobacco shred,and K‑means clustering is performed on the processed images to obtain the foreground and hindground of images and complete segmentation to improve the environmental interference resistance and feature extraction capability of the input images. An efficient channel attention mechanism is introduced into the Inception‑ResNet‑V2 network to strengthen the model’s ability to extract features;at the same time,the improved module outputs feature maps for multi‑scale fusion to increase feature representation and reduce the risk of overfitting. Finally,the ReLU activation function and Adam optimizer are replaced by PReLU and AdaBound when comparing convergence and accuracy.The experimental results showed that the proposed algorithm had good generalization ability and enabled efficient identification of four types of tobacco shred,with final recognition accuracy of 97.23% and detection time of 0.107 s for a single image.

Key words: Tobacco shred, K?means algorithm, Inception network, Efficient channel attention mechanism, Multi?scale feature fusion

中图分类号: