高精度轻量级烟叶烘烤阶段识别模型研究  

Research on high-precision lightweight model for tobacco leaf curing stage identification

作　　者：姜增昀 祝诗平[1] 冯川 王涛 李俊贤 JIANG Zengyun;ZHU Shiping;FENG Chuan;WANG Tao;LI Junxian(College of engineering and technology,Southwest University,Chongqing 400715,China;Qujing City Tobacco Company,Qujing 655000,Yunnan,China)

机构地区：[1]西南大学工程技术学院,重庆市400715 [2]云南省曲靖烟草公司,云南省曲靖市655000

出　　处：《中国烟草学报》2023年第1期55-63,共9页Acta Tabacaria Sinica

基　　金：中国烟草总公司云南省公司科技重大项目(2021530000241036)。

摘　　要：【背景】实时的烟叶状态识别和烘烤阶段判别是实现烘烤工艺精准调节和建立烟叶智能烘烤系统的关键技术。目前烟叶烘烤状态识别方面的研究受到模型规模的制约,难以应用于实际生产,因此有必要研究高精度轻型识别模型。【过程】采用工业摄像头和补光灯组成的图像采集系统采集烘烤过程的烟叶图像,通过伽马变换和HSI颜色空间转化并对烟叶数据集进行处理,使用改进的轻量级EfficientNetB0模型进行训练,并在Jetson Nano嵌入式开发板上进行模型可行性测试。【结果】改进模型在测试集上的准确率达到了96.13%,参数量仅为2.74 M,相比于原始EfficientNetB0,改进模型的识别准确率提高了1.59%,参数量减少了48.50%。在Jetson Nano开发板上,相比于MobileNetv2、MobieleNetv3等轻型模型,改进模型的识别速度和准确率均有明显提升。【结论】改进的EfficientNetB0模型可以实现烟叶烘烤阶段的快速、精准识别。同时本研究可为烟叶智能烘烤系统的建立提供了理论依据和现实参考。[Background] Real-time identification of tobacco leaf status and its curing stage is key a technology to achieve accurate adjustment of tobacco curing process and establish an intelligent tobacco leaf curing system. At present, the research on identification of tobacco curing status is restricted by the scale of the model, and it is difficult to apply current models to actual production. Therefore, it is necessary to develop a high-precision lightweight model for tobacco leaf curing stage identification. [Methods] An image acquisition system consisting of an industrial camera and a fill light was used to collect tobacco images during the curing process. The tobacco dataset was processed through gamma transformation and HSI color space transformation. The improved lightweight EfficientNetB0 model was used for training, and the model feasibility test was conducted on the Jetson Nano embedded development board. [Results] The accuracy of the improved model in the test set reached 96.13%, and the parameter quantity was only 2.74M. Compared with the original Efficient NetB0, the identification accuracy of the improved model was increased by 1.59%, and the parameter quantity was reduced by 48.50%. On the Jetson Nano development board, compared with the lightweight models such as MobileNetv2 and MobileNetv3, the recognition speed and accuracy of the improved model were significantly improved. [Conclusion] The improved Efficient NetB0 model can realize fast and accurate identification of tobacco leaf curing stage. This study provides a theoretical basis and practical reference for the establishment of intelligent tobacco curing system.

关 键 词：烤烟阶段识别 EfficientNet 伽马变换 HSI颜色空间 ECA注意力机制 

分 类 号：TS441[农业科学—烟草工业]