基于渐进式学习和增强原型度量的小样本农作物病害识别方法

Few-shot Crop Disease Recognition Based on Progressive Learning and Enhanced Prototype Metric

摘要:

为了开展低成本、通用、灵活的农作物病害识别，提出了一种基于渐进式学习和增强原型度量的小样本农作物病害识别网络(Few-shot crop disease recognition network based on progressive learning and enhanced prototype metric, FPE-Net)。首先，利用设计的增强原型度量模块，计算能够准确表示类别中心的增强原型，并充分利用增强原型中的丰富类别信息对农作物病害进行识别；其次，采用设计的渐进式学习策略对模型进行训练，以帮助模型更好地适应农作物病害识别任务，提升模型小样本农作物病害识别精度。在自制小样本农作物病害数据集FSCD-Base、FSCD-Complex以及FSCD-Base到FSCD-Complex的跨域设置上，FPE-Net的5-way 1-shot平均识别准确率分别达到70.65%、53.47%和49.58%，5-way 5-shot平均识别准确率分别达到83.02%、66.15%和64.21%。实验结果表明，本文提出的FPE-Net明显优于其他小样本农作物病害识别模型，在训练样本不足的情况下能够更准确识别农作物病害。

Abstract:

At present, crop disease recognition is mostly realized based on convolutional neural network. However, due to the lack of training data in actual agricultural production, these crop disease recognition methods based on convolutional neural network often have limited applications and perform poorly. In order to carry out the low-cost, general and flexible crop disease recognition, a fewshot crop disease recognition network based on progressive learning and enhanced prototype metric was proposed. Specifically, an enhanced prototype metric module was firstly designed to compute the enhanced prototype that can accurately represent the category center, and make full use of its rich category information to recognize the crop disease. Then, a progressive learning strategy was designed to train the model to help it better adapt to the crop disease recognition, and further improve the few-shot crop disease recognition accuracy. On the self-made few-shot crop disease datasets FSCD-Base, FSCD-Complex and the cross-domain setting from FSCD-Base to FSCD-Complex, the 5-way 1-shot average recognition accuracy of the FPE-Net reached 70.65%, 53.47% and 49.58%, and the 5-way 5-shot average recognition accuracy of the FPE-Net reached 83.02%, 66.15% and 64.21%, respectively. These experimental results showed that the FPE-Net was significantly better than other fewshot crop disease recognition models, which can recognize crop diseases more accurately when the training data was insufficient.