【摘要】 针对面对害虫数据稀缺的实际场景时，现有害虫图像识别方法容易出现过拟合导致模型表达能力不足的问题，本研究提出了一种结合度量学习和迁移学习的小样本田间害虫图像分类识别方法。首先，使用ECA-Pyramid-ResNet12模型在mini-ImageNet数据集上进行预训练；其次，在度量模块中添加ECA通道注意力机制，通过捕捉通道间的依赖关系来增强害虫的图像特征表示；然后，使用特征金字塔结构来捕获害虫图像的局部特征和害虫的多尺度特征；最后，利用20类自建鳞翅目害虫图像作为元数据集，对模型进行元训练和元测试。实验结果表明，在3-way 5-shot和5-way 5-shot条件下，本文模型准确率分别达到91.16%和87.26%,比SSFormers、DeepBDC方法分别提高4.58、1.35个百分点。提出的模型有效提升了小样本学习中目标图像特征的表达能力，能够为数据稀缺场景下的田间害虫自动识别提供方法参考。更多还原

【Abstract】 In real-world scenarios where pest data is scarce, existing pest image recognition methods are prone to overfitting, resulting in insufficient model expressiveness. To address this issue, a novel few-shot field pest image classification method that integrated metric learning with transfer learning was proposed. Firstly, the ECA-Pyramid-ResNet12 model was pretrained on the mini-ImageNet dataset. Subsequently, PN was chosen as the classifier, and cosine similarity was selected as the distance metric. The ECA channel attention mechanism was then incorporated into the metric module to enhance pest image feature representation by capturing inter-channel dependencies, with a kernel size of 3. Additionally, a feature pyramid structure was employed to capture the local and multi-scale features of pest images. After evaluating different pooling combinations, the 2×2+4×4 pooling combination was selected. Finally, a meta-dataset comprising 20 self-built categories of Lepidoptera pest images was utilized for meta-training and meta-testing of the model. Experimental results demonstrated that under 3-way 5-shot and 5-way 5-shot conditions, the proposed method achieved accuracy rates of 91.16% and 87.26%, respectively, surpassing the most relevant works of the past two years, SSFormers and DeepBDC, by 4.58 percentage points and 1.35 percentage points. The proposed model effectively enhanced the feature representation of target images in few-shot learning, providing a methodological reference for the automatic identification of field pests in data-scarce scenarios.更多还原