摘要
为了解决小麦种植中品系混乱、劣种降效、假种坑农以及模型参数过多不利于部署到移动端等问题,提出了IRMAnet模型。通过拍摄29种不同小麦的种子期、幼苗期、开花期图片,构建了一个拥有87个类别46 420张照片的小麦多生育时期数据集。基于该数据集,首先将原始ResNet34模型的基本残差块中的第二个卷积块替换为inverted residual block,以降低网络的参数量;其次在网络的Layer1层后加入一层RFB层,增大感受野的同时提高特征提取能力;最后在网络的Layer2、Layer3层后分别加入一层MAPOOL层,以增强泛化能力和准确性。在训练集上进行训练后,IRMAnet的准确率为95.0%,相较于ResNet34提高了1.9个百分点。将在训练集上训练得到的权重加载到验证集上后,除个别品种外,绝大多数品种的精确率、召回率、特异度均达到了90%以上。实验结果表明,IRMAnet能够对多个生育时期的小麦品种进行准确识别,模型性能更加优越,所使用参数量更低。该研究为全生育期小麦品种识别提供了依据,为小麦产业提质增效提供了新的技术选项。
In order to solve some issues in wheat cultivation,including strain confusion,poor seeds decreasing efficiency,fake seeds deceiving farmers,and excessive model parameters hindering mobile device deployment,the IRMAnet model was proposed.A wheat variety in the whole growth period dataset with 46420 photos of 87 categories was constructed by taking pictures of 29 different wheat varieties at seeding,seedling,and flowering stages.Based on the dataset,the basic residual blocks in the original ResNet34 model were modified by replacing the second convolutional block with an Inverted Residual Block,reducing network parameters.Additionally,a RFB layer was introduced after Layer1 to enlarge receptive fields and enhance feature extraction,followed by the incorporation of MAPOOL layers after Layer2 and Layer3 to improve generalization and accuracy.After training on the training set,the accuracy of IRMAnet is 95.0%,with an increase rate of 1.9 percentage points compared to ResNet34.Upon applying the trained weights to the validation set,most wheat varieties exhibited precision,recall,and specificity exceeding 90%.The experimental results show that IRMAnet is able to accurately identify wheat varieties at multiple growth stages,with more superior model performance and lower number of parameters used.This study provides a foundation for identifying wheat varieties throughout their entire growth cycle,offering new technological options to enhance the quality and productivity of the wheat industry.
作者
冯永强
刘成忠
韩俊英
鲁清林
刘立群
邢雪
FENG Yongqiang;LIU Chengzhong;HAN Junying;LU Qinglin;LIU Liqun;XING Xue(College of Information Sciences and Technology,Gansu Agricultural University,Lanzhou,Gansu 730070,China;Wheat Research Institute,Gansu Academy of Agricultural Sciences,Lanzhou,Gansu 730070,China)
出处
《麦类作物学报》
CAS
CSCD
北大核心
2024年第2期242-252,共11页
Journal of Triticeae Crops
基金
甘肃省高等学校创新基金项目(2021A-056)
甘肃省高等学校产业支撑计划项目(2021CYZC-57)
国家自然科学基金项目(32160421)。
