摘要
红火蚁是近年来侵害我国南方的主要外来入侵物种之一,精确识别红火蚁巢穴是防控红火蚁的关键所在。为解决传统红火蚁防控依赖人工巡视、高危险、效率低的问题,降本增效实现红火蚁巢穴的智能检视,提出了一种基于ResNet34改进的红火蚁巢穴判别模型。该模型借助采集于不同地貌特征下的红火蚁巢穴图像,结合数据增强技术进行训练,通过在ResNet34的第1层卷积层之后和全连接层之前加入SE注意力机制模块,提升网络的自适应选择和通道权值调整能力,以提取红火蚁巢穴表面局部非线性的纹理特征。经过K折交叉验证试验和超参数探究消融试验,将SE-ResNet34与AlexNet、VGG-16、ResNet18、ResNet34、ResNet50进行对比,分析得出SE-ResNet34的峰值准确率达到了98.76%,比ResNet34的准确率提高了2.17%,较其他测试模型有训练时间短、识别精度高的特点,同时展现出较强的鲁棒性和稳定性。该方法在减少人工成本的同时可降低杀虫剂的使用,为红火蚁巢穴判别提供了一种便捷高效的解决方案。
The red imported fire ant(solenopsis invicta buren)is one of the major invasive species that has been causing damage in southern China in recent years.Accurately identifying the nests of red imported fire ants is crucial for their prevention and control.In this paper,an improved red fire ant nest identification model based on ResNet34 is proposed to solve the problems of traditional red fire ant prevention and control relying on manual inspection,high risk and low efficiency,so as to realize intelligent inspection.This model uses red imported fire ant nest images collected from different terrain features,combined with data augmentation techniques for training.By adding an SE attention mechanism module after the first convolutional layer of ResNet34 and before the fully connected layer,the model can enhance its adaptive selection and channel weighting adjustment capabilities to extract local nonlinear texture features on the surface of red imported fire ant nests.Through K-fold cross-validation tests and ablation tests to explore hyperparameters,SE-ResNet34 is compared with AlexNet,VGG-16,ResNet18,ResNet34,and ResNet50,and the results show that SE-ResNet34 achieves a peak accuracy of 98.76%,which is 2.17%higher than ResNet34.It has the characteristics of short training time,high recognition accuracy,strong robustness,and stability compared to other tested models.This method provides a convenient and efficient solution for distinguishing red fire ant nests while reducing manual labor and minimizing the use of insecticides.
作者
袁嘉辉
刘蕊
梁虹
周祥
Yuan Jiahui;Liu Rui;Liang Hong;Zhou Xiang(School of Information,Yunnan University,Kunming 650504,China;School of Earth Sciences,Yunnan University,Kunming 650504,China;School of Plant Protection,Hainan University,Haikou 570228,China)
出处
《电子测量技术》
北大核心
2023年第23期97-104,共8页
Electronic Measurement Technology
基金
云南大学第二届专业学位研究生实践创新项目(ZC-22222790)资助
