English

Journal of Chinese Agricultural Mechanization

Journal of Chinese Agricultural Mechanization ›› 2025, Vol. 46 ›› Issue (2): 120-125.DOI: 10.13733/j.jcam.issn.2095‑5553.2025.02.018

Previous Articles     Next Articles

A lightweight shiitake quality classification model based on improved VGG

Sun Yan, Zhu Fengwu, Zhang Yuqing, Zhang Weijian, Wu Chengxuan   

  • Online:2025-02-15 Published:2025-01-24

基于改进VGG的轻量级香菇品质分类模型 

孙岩,朱凤武,张宇清,张伟健,吴铖轩   

  • 基金资助:
    国家重点研发计划(2020YFD1000304—5);吉林省科技发展计划项目(20210202054NC)

Abstract: In order to achieve rapid and accurate classification of shiitake mushroom quality, according to the characteristics of shiitake mushroom images in greenhouse planting environment, a shiitake quality classification detection model is proposed based on a lightweight improvement of VGG16 deep convolutional network. Firstly, the VGG16 network is lightweighted, and the hybrid pooling layer is used instead of the fully connected layer to downsample the feature map. Then, the hole fusion separation convolution and channel attention SE modules are introduced in the feature extraction network to improve the recognition accuracy of the model. After that, the dataset is enriched with data augmentation methods. Finally, transfer learning is used to train the shiitake mushroom quality classification detection model. Under the same experimental conditions, five deep convolutional network models are compared with VGG16, GoogLeNet, VGG19, ResNet50 and MobileNetv1. The results show that the comprehensive performance of the proposed model is the best, and the recognition accuracy of the improved VGG16 network is 95.5%. The model size is about 10.9% of the original VGG16 model volume. The training time is 55.1% of the original VGG16 model. 

Key words: Lentinula edodes, shiitake quality classification, VGG16, SE module, equalization pooling, void fusion separation convolution

摘要: 为实现对香菇品质的快速、准确分类,依据大棚种植环境下香菇图像的特性,对VGG16深度卷积网络进行轻量化改进,提出一种香菇品质分类检测模型。首先,对VGG16网络进行轻量化处理,利用均和池化层代替全连接层对特征图进行下采样;然后,在特征提取网络中引入空洞融合分离卷积和通道注意力SE模块提升模型的识别精度;之后,利用数据增强方法将数据集扩充;最后,使用迁移学习训练得到香菇品质分类检测模型。在相同的试验条件下,与VGG16、GoogLeNet、VGG19、ResNet50、MobileNetv1五种深度卷积网络模型相比较。结果表明:该模型的综合性能最好,改进后的VGG16网络的识别准确率为95.5%;模型大小约为原始VGG16模型体量的10.9%;训练时间为原始VGG16模型的55.1%。

关键词: 香菇, 品质分类, VGG16, SE模块, 均和池化, 空洞融合分离卷积

CLC Number: