| 摘要: |
| 【目的】针对自然状态烟叶传统正副组分类速度慢、正副组易错分、特征提取困难的问题，提出了一种基于轻量级SE PPM的自然状态烟叶正副组分类算法（SAPMDSNet）。【方法】基于轻量级ShuffleNetV2网络，先通过降低网络卷积深度和进化激活函数，加快网络模型的训练速度；再引入通道注意力机制SE模块，增强通道间的特征差异，提高网络模型的表征能力，避免正副组烟叶叶部区域化导致的组别错分；最后通过嵌入金字塔池化模块PPM充分融合烟叶显露特征与全局信息，增强对正副组烟叶上下文信息的聚合,并采用自行构建的烟叶数据集进行对比试验。【结果】SAPMDSNet网络模型的分类准确率为91.09%，计算量（FLOPs）为151.70 M，取得了相对较高的分类效果。与原网络ShuffleNetV2模型和轻量级GhostNet模型相比，SAPMDSNet网络模型的FLOPs分别升高2.65%和2.84%，而识别准确率则分别提高2.72和21.13个百分点；MobileNetV2、DenseNet和SqueezeNet模型的识别准确率分别为87.02%，89.53%和87.60%，虽均与SAPMDSNet模型的识别准确率接近，但其FLOPs明显较SAPMDSNet模型大。【结论】构建的SAPMDSNet模型能提高烟叶正副组分类精度且具有较好的整体性能，为烤烟烟叶品质初筛提供了新的思路和方法。 |
| 关键词: 自然状态烟叶 正副组分类 轻量化模型 注意力机制 金字塔池化 |
| DOI： |
| 分类号: |
| 基金项目:云南省重大科技专项(202002AD080001) |
|
| Classification algorithm of positive and sub-group tobacco leaves in natural state based on lightweight SE-PPM |
|
WANG Hongcheng,GU Wenjuan,LIU Xiaobao,et al
|
| Abstract: |
| 【Objective】A lightweight SE-PPM algorithm (SAPMDSNet) was proposed to solve the problems of slow speed,misclassification and difficulty in feature extraction in traditional positive and sub-group classification of natural state tobacco leaves.【Method】Based on the lightweight ShuffleNetV2 network,the training speed of the network model was accelerated by reducing the network convolution depth and evolving the activation function.Then,the channel attention SE module was introduced to enhance the characteristic differences between channels,improve the representation ability of the network model and avoid the group misclassification caused by the regionalization.Finally,the pyramid pooling module PPM was embedded to integrate the exposure characters and global information by aggregating context information,and the in-house tobacco leaf data set was used and compared with other models.【Result】The SAPMDSNet network model achieved relatively high classification results with accuracy of 91.09% and FLOPs of 151.70 M.Compared with the original ShuffleNetV2 model and the lightweight GhostNet model,the FLOPs were slightly increased by 2.65% and 2.84%,and the accuracy was increased by 2.72% and 21.13%,respectively.MobileNetV2,DenseNet and SqueezeNet achieved recognition accuracies of 87.02%, 89.53% and 87.60%,which were close to the proposed model,but their FLOPs were significantly higher.【Conclusion】The constructed SAPMDSNet network model improved the classification accuracy of positive and sub-group tobacco leaves in natural state with better overall performance,which provides a new method for primary screening of tobacco leaves. |
| Key words: natural state group tobacco leaves positive and sub-group classification lightweight model attention mechanism pyramid pooling |
