摘要
【目的】为提高树种识别工作的效率和准确率,提出一种利用迁移学习策略并引入SimAM注意力机制和ECA通道注意力机制的ConvNeXt树种识别模型。【方法】以12种常见树种的树皮图像为研究对象,通过传统数据增强方法对数据进行扩充,防止模型过拟合。使用SimAM和ECA通道注意力机制构建以ConvNeXt为基础的改进网络,增强特征提取的SA-ConvNeXt、增强重要特征权重的E-ConvNeXt、结合两者的ES-ConvNeXt,测试数据集在增强前后对ES-ConvNeXt网络准确率的影响。使用ResNet34、ResNet50、GoogLeNet、Swin Transformer、DenseNet121和ConvNeXt网络,与ES-ConvNeXt模型识别效果进行比较。【结果】SA-ConvNeXt和E-ConvNeXt准确率分别达到(95.14±0.42)%、(96.085±0.235)%,ES-ConvNeXt在增强后数据集测试的准确率达到(97.445±0.635)%,对单一树种识别准确率均超过93%,最高类别准确率达到99.79%,为最优方案。经数据增强后进行训练的模型与使用原始数据进行训练的模型相比,其验证集的准确率和损失值,无论是收敛速度还是最终稳定值都是最优。数据集相同时,使用ResNet34、ResNet50、GoogLeNet、Swin Transformer、DenseNet121和ConvNeXt网络的识别准确率,分别为92.74%、94.47%、90.52%、92.85%、70.38%、94.72%,均低于新改进模型ES-ConvNeXt(97.81%),进一步说明了改进后的ESConvNeXt模型的有效性。【结论】数据增强对模型准确率提升有效,在数据增强后的数据集上,改进后的ESConvNeXt模型与其他模型相比可以更加准确地完成树种分类任务,在不同树种上也有较好的泛化能力。
【Objective】In this study,an improved tree species recognition model of ConvNeXt network was proposed by using a transfer learning strategy and introducing the SimAM attention module and ECA channel attention mechanism,so as to improve the efficiency and accuracy of tree species recognition work and solve the difficulties encountered in the recognition work.【Method】The bark images of common 12 tree species were used as the research object,and the data were expanded by traditional data enhancement methods to prevent model overfitting.An improved ConvNeXt-based network was constructed using SimAM and ECA channel attention mechanisms:SA-ConvNeXt for enhanced feature extraction,E-ConvNeXt for enhanced weighting of important features,and ES-ConvNeXt combining the two.The effect of the dataset on the accuracy of the ESConvNeXt network before and after enhancement was tested.The recognition effects with the ES-ConvNeXt model were compared by using the Resnet34,Rennet50,GoogLeNet,Swin Transformer,Densenet121,and ConvNeXt networks.【Result】SA-ConvNeXt and E-ConvNeXt achieved 95.14%±0.42%and 96.085%±0.235%accuracy,respectively.ES-ConvNeXt,which incorporates SimAm and ECA attention modules,achieved an accuracy of 97.445%±0.635%for the test on the augmented dataset,its recognition accuracy for a single tree species exceeded 93%,and the highest category accuracy reached 99.79%,making it the optimal solution.The model trained with expanded data had optimal accuracy and loss values for the validation set both in terms of speed of convergence and final stabilized values compared to the model trained using the original data.With the same dataset,the recognition accuracies using Resnet34,Rennet50,GoogLeNet,Swin Transformer,Densenet121,and ConvNeXt networks were 92.74%,94.47%,90.52%,92.85%,70.38%,and 94.72%,respectively,which were all lower than the 97.81%obtained by the new improved model(ES-ConvNeXt model),further illustrating the effectiveness of the improved ES-ConvNeXt model.【Conclusion】Data enhancement is effective for model accuracy improvement,and on the data-enhanced dataset,the improved ES-ConvNeXt model can perform the tree classification task more accurately compared to the other models,and it also has better generalization ability on different tree species.
作者
杨兵兵
许杰
Yang Bingbing;Xu Jie(College of Information and Electrical Engineering,Heilongjiang Bayi Agricultural University Daqing 163319)
出处
《林业科学》
北大核心
2025年第2期31-39,共9页
Scientia Silvae Sinicae
基金
国家自然科学基金项目(31570712)
黑龙江省自然基金项目(LH2022E099)。
