基于MobileNetV2和卷积注意力机制的轻量化玉米籽粒品种识别研究  

Lightweight Maize Seed Variety Recognition Model Based on MobileNetV2 and Convolution Attention Mechanism

作　　者：孙孟研 孙彤辉 郝凤琦 穆春华[5] 马德新[1] Sun Mengyan;Sun Tonghui;Hao Fengqi;Mu Chunhua;Ma Dexin(College of Communication,Qingdao Agricultural University,Qingdao 266109,China;Shandong Cultural Industry Vocational College,Qingdao 266699,China;College of Computer Science and Technology,Qingdao University,Qingdao 266071,China;Shandong Computer Science Center(National Supercomputer Center in Jinan),Jinan 250014,China;Maize Research Institute,Shandong Academy of Agricultural Sciences,Jinan 250100,China)

机构地区：[1]青岛农业大学传媒学院,山东青岛266109 [2]山东文化产业职业学院,山东青岛266699 [3]青岛大学计算机科学技术学院,山东青岛266071 [4]山东省计算中心(国家超级计算济南中心),山东济南250014 [5]山东省农业科学院玉米研究所,山东济南250100

出　　处：《山东农业科学》2024年第12期139-146,共8页Shandong Agricultural Sciences

基　　金：山东省自然科学基金项目(ZR2022MC152);中央引导地方科技发展专项计划(23-1-3-6-zyyd-nsh);山东省重点研发计划项目(2023TZXD023)。

摘　　要：快速、准确地识别农作物品种对我国粮食安全和农业发展具有重要意义。为实现玉米种子的快速鉴别与保护,本研究提出一种基于MobileNetV2和卷积注意力机制的玉米籽粒品种识别算法。首先购得市面上9个常规玉米品种的籽粒,使用佳能80D型相机对其胚面和胚乳面进行图像采集,构建了包含3408张图像的玉米籽粒识别数据集,按照7∶2∶1划分训练集、验证集和测试集,并对训练集图像进行数据增强处理;然后设计注意力模块ISPAM(Improved Spatial Attention Module),即在卷积注意力模块(CBAM)基础上,提出一种新的通道注意力模块ICAM对CBAM的通道注意力机制进行改进,同时引入空间金字塔池化(SPP)模块替换CBAM空间注意力模块中的平均池化模块和最大池化模块,构建了玉米籽粒品种识别模型MobileNetV2_ISPAM。将MobileNetV2_ISPAM与添加其他注意力模块的模型对比,结果表明,MobileNetV2_ISPAM在测试集上的准确率为99.11%,均明显高于MobileNetV2以及添加SE(Squeeze-and-Excitation)、CBAM注意力机制的模型。梯度加权类激活映射网络可视化表明,MobileNetV2_ISPAM更关注玉米籽粒图像中的显著特征,从而提高了模型的准确率。此外,该模型的参数量仅为7.15 M,适合移动端的便携化部署。本研究在保证模型轻量高效的前提下,提升其抵抗过拟合能力和分类性能,为以后基于深度学习的移动端玉米籽粒图像识别模型研究提供了思路。Rapid and accurate identification of crop varieties is of great significance for ensuring food security and improving the quality of agricultural development in our country.In order to achieve fast identification and protection of maize seeds,this experiment proposed a maize seed variety recognition algorithm based on MobileNetV2 and attention mechanism.Firstly,nine popular maize seed varieties were obtained from the market,and their images were captured using a Canon 80D camera to create a dataset of 3408 maize seed images.The dataset was divided into training,validation,and testing sets in a ratio of 7:2:1 and subjected to data augmentation.Then,an attention module called ISPAM(Improved Spatial Attention Module)and a maize seed variety recognition model called MobileNetV2_ISPAM(MobileNetV2_Improved Spatial Attention Module)were designed.Finally,MobileNetV2_ISPAM was compared with models incorporating other attention modules.The experimental results show that MobileNetV2_ISPAM achieved an accuracy of 99.11%on the test set,which is an improvement of 12.11%,6.8%,and 2.36%compared to MobileNetV2,models with SE(Squeeze-and-Excitation)and CBAM attention mechanisms,respectively.The visualization of the gradient-weighted class activation mapping network demonstrated that MobileNetV2_ISPAM payed more attention to salient features in maize seed images,thus improving the accuracy of the model.Moreover,the model has a parameter size of only 7.15MB,making it suitable for deployment on mobile devices.This experiment enhanced the ability to resist overfitting and improved classification performance while ensuring a lightweight and efficient model,providing insights for future research on deep learning-based maize seed image recognition models for mobile platforms.

关 键 词：MobileNetV2 ISPAM注意力机制 深度学习 玉米籽粒 品种识别 

分 类 号：S126[农业科学—农业基础科学]