农业灯诱害虫图像识别的模型算法研究

Research on Insect Identification Model Algorithm for AgriculturalLamp-baiting Images

【目的】采用Faster R-CNN算法对样本数量少且分布不均衡的28类农田害虫数据集进行研究。【方法】首先,分析不同输入图像尺寸对训练模型性能的影响,确定了输入图像尺寸5472×3648的25%作为优选;其次,为了避免部分类别害虫因数据过少而导致的过拟合问题,采用Mixup和mosaic方法增加数据多样性,并使用迁移学习提高模型性能。【结果】这些方法可以有效地提高模型的泛化性和鲁棒性,除了9与10这两类害虫相似度非常高导致AP值较低外,其余害虫识别的AP平均值为92.07%。【结论】通过测试数据检验模型的泛化性,发现模型表现良好但仍有改进空间。

[Purposes]This study aims to identify 28 types of farmland pests with small samples and un-balanced datasets using Faster R-CNN algorithm.[Methods]Firstly,the impact of different input image sizes on the performance of the training model is analyzed,and the optimal selection is determined as 25%of the input image size.Secondly,in order to avoid overfitting problems caused by too few pest data for some classifications,Mixup and Mosaic methods are used to increase data diversity,and transfer learning is used to improve the model performance.[Findings]These methods can effectively improve the generalization ability and robustness of the model.Except for the two kinds of pests No.9 and No.10,which have very high similarity and low AP value,the average AP value of other pests reaches 92.07%.[Conclusions]The generalization ability of the model is verified by the test data.The model performs well but still has room for improvement.