基于专家模型混合的细粒度烟叶分级方法-《计算机技术与发展》

文章信息/Info

Title:: Fine-grained Tobacco Leaf Grading Method Based on Expert Model Mixture

作者:: 权传明1; 陈烨1; 2; 梅豪1; 1. 南京工程学院人工智能产业技术研究院,江苏南京 211167; 2. 江苏省智能感知技术与装备工程研究中心,江苏南京 211167

Author(s):: QUAN Chuan-ming1; CHEN Ye1; 2; MEI Hao1; 1. AI Industrial Technology Research Institute,Nanjing Institute of Technology,Nanjing 211167,China; 2. Jiangsu Intelligent Perception Technology and Equipment Engineering Research Center,Nanjing 211167,China

Keywords:: tobacco leaf grading; expert mode; deep learning; fine-grained grading; attention mechanism

摘要:: 针对卷烟生产烟叶分级过程中人工劳动强度大、易受主观因素影响等问题,提出一种基于细粒度专家混合模型的烟叶分级方法。该方法基于三级专家模型架构,通过下级专家模型不断学习凝练上级专家模型给出的先验信息及图像特征从而增强模型特征表征能力及烟叶分级精度。针对烟叶数据集类别差异不明显导致不同专家模型之间易产生相似性预测的问题,提出一种基于MGE_CNN专家模型混合的细粒度烟叶分级方法,并且引入KL散度约束和PReLU激活函数进一步提升模型训练收敛效果。经实际卷烟生产烟叶分级图像数据集测试,该方法在实际生产环境中的烟叶分级精度可达90.47%,相较于现有细粒度分级方法精度提升0.95%,达到卷烟生产烟叶分级精度指标要求,可为后续实现自动化烟叶分级提供有力支撑。

Abstract:: In response to the issues of high manual labor intensity and susceptibility to subjective factors in the tobacco leaf grading process during cigarette production,a tobacco leaf grading method based on a fine-grained expert mixture model is proposed. The method is built on a three-tier expert model architecture,where the lower-level expert models continuously learn and refine the prior in-formation and image features provided by the upper-level expert models,thereby enhancing the model's feature representation ability and improving the tobacco leaf grading accuracy. To address the problem of similar predictions between different expert models caused by the subtle differences in categories within the tobacco leaf dataset,a fine-grained tobacco leaf grading method based on the MGE_CNN expert model mixture is introduced. Additionally,KL divergence constraints and the PReLU activation function are incorporated to further improve the model's convergence during training. Testing on an actual tobacco leaf grading image dataset from the cigarette production process shows that the proposed method achieves a grading accuracy of 90.47%,which is an improvement of 0.95% compared to existing fine-grained grading methods. This meets the tobacco leaf grading accuracy requirements for cigarette production and provides strong support for the subsequent realization of automated tobacco leaf grading.