基于改进YOLOv3的玉米病害识别方法  

Maize disease identification method based on improved YOLOv3

作　　者：张继成 黄向党[2] Zhang Jicheng;Huang Xiangdang(Jingzhou University,Jingzhou,434020,China;School of Computer Science and Technology,Hainan University,Haikou,570228,China)

机构地区：[1]荆州学院,湖北荆州434020 [2]海南大学计算机科学与技术学院,海口市570228

出　　处：《中国农机化学报》2024年第7期269-275,共7页Journal of Chinese Agricultural Mechanization

基　　金：湖北省教育厅科学研究计划项目(B2020340);长江大学工程技术学院(荆州学院)科学研究基金项目(2020KY04)。

摘　　要：为提高玉米作物病害叶片识别模型的准确性,提出改进YOLOv3的玉米病害识别方法。首先,为获得更深的玉米疾病特征,通过更改浅特征图比例和添加第四个检测层,分别修改YOLOv3网络体系结构为YOLOv3-M1和YOLOv3-M2。然后,采用改进的K-means算法进行聚类,获得的锚框倾向于数据集的真实边界框。最后,为每个类别添加一个平衡因子,并对不同类别中样本的难度进行加权来修改损失函数,使得模型能够找到边界盒预测与类别预测之间的最佳点,使算法获得最佳检测效果。结果表明,改进的YOLOv3-M1和YOLOv3-M2模型在测试集上的准确率分别高达95.63%和97.59%,相比YOLOv3模型,识别准确率分别提高4.15%和6.28%,识别准确率在玉米数据集上得到大幅度提高。In order to improve the accuracy of maize disease leaf recognition model,an improved YOLOv3 maize disease recognition method was proposed.First of all,in order to obtain deeper maize disease characteristics,the YOLOv3 network architecture was modified to YOLOv3-M1 and YOLOv3-M2 by changing the proportion of shallow feature map and adding a fourth detection layer.Then,the improved K-means algorithm was used for clustering,and the obtained anchor frame tended to be the true boundary frame of the data set.Finally,a balance factor was added for each category,and the difficulty of samples in different categories was weighted to modify the loss function,so that the model could find the best point between the boundary box prediction and the category prediction,so that the algorithm could obtain the best detection effect.The test results show that the accuracy of the improved YOLOv3-M1 and YOLOv3-M2 models in the test set is as high as 95.63%and 97.59%,respectively.Compared with the YOLOv3 model,the recognition accuracy is increased by 4.15%and 6.28%,respectively,and the recognition accuracy is greatly improved in the corn data set.

关 键 词：玉米 深度学习 病害识别 YOLOv3模型 损失函数 

分 类 号：S24[农业科学—农业电气化与自动化] TP391.4[农业科学—农业工程]