摘要
为提升草地贪夜蛾及其近缘种成虫识别模型的泛化能力,除识别准确率外,额外引入特征识别率对模型的泛化能力进行评估。将VGG-16-bn模型的全连接层以全局平均值池化层取代,并在模型训练阶段引入了Grad-CAM可视化结果进行训练指导,共构建了4种改进模型识别草地贪夜蛾及其近缘种成虫。结果表明,改进后的模型的识别准确率均在99.22%以上,VGG-16-bn-GAP模型参数内存需求仅为原始模型的10.98%。为评估模型的泛化能力,利用导向反向传播梯度值、Grad-CAM及Grad-CAM++对模型习得的特征进行可视化,并与专家进行人工识别的关键视觉特征进行比较。结果表明,改进的VGG-16-bn-GAP模型和VGG-16-bn-GAIN模型获得的草地贪夜蛾平均特征识别率比原始模型分别提高12.25%和13.42%。本文提出的以特征识别率评估模型泛化能力的方法,可为特征识别率和识别准确率的提升提供参考。
In order to improve the generalization ability of the recognition model for Spodoptera frugiperda and its related species adult,in addition to recognition accuracy,an additional feature recognition rate was introduced to evaluate the generalization ability of the model.The fully connected layer of the VGG-16-bn model was replaced by a global average pooling layer,and Grad-CAM visualization results were introduced for training guidance in the model training phase.A total of four improved models were constructed to recognize the Spodoptera frugiperda and its related species adult.The results showed that the recognition accuracy of the improved model was above 99.22%,and the memory requirement of the VGG-16-bn-GAP model parameters was only 10.98%of the original model.In order to evaluate the generalization ability of the model,guided backpropagation gradient values,Grad-CAM,and Grad-CAM++were used to visualize the features learned by the model,and compared with key visual features manually recognized by experts.The average feature recognition rate of Spodoptera frugiperda obtained by the improved VGG-16-bn-GAP model and VGG-16-bn-GAIN model was 12.25%and 13.42%higher than that of the original model.The method for evaluating the generalization ability of the model by feature recognition rate proposed in this paper can provide references for the improvement of feature recognition rate and recognition accuracy.
作者
魏靖
季英超
WEI Jing;JI Yingchao(Shenzhen Agricultural Science Group Co.,Ltd.,Shenzhen Guangdong 518000;School of Plant Protection,Shandong Agricultural University,Taian Shandong 271000)
出处
《现代农业科技》
2024年第8期163-169,共7页
Modern Agricultural Science and Technology
