基于多级Logit知识蒸馏的植物病害识别研究

Plant Disease Recognition Based on Multi-Level Logit Knowledge Distillation

摘要: 植物病害是威胁全球粮食安全的重要因素，传统人工检测方式难以满足现代智慧农业对精准化、自动化病害识别的需求。深度学习虽为此提供了解决思路，但常规卷积神经网络参数量大、计算复杂度高，而轻量化网络又面临分类精度不足的问题。为此，本文提出一种基于多级Logit知识蒸馏的植物病害分类方法。该方法以ResNet34为教师模型，通过联合实例级、批次级与类别级蒸馏损失，从多个维度将教师网络的知识迁移至轻量学生模型中，在不增加推理开销的前提下提升其分类性能。基于Kaggle公开的多源植物病害分类综合数据集进行实验。结果表明，蒸馏后学生模型的Top-1准确率最高可达95.19%，超越教师模型的94.20%。CAM可视化进一步验证了蒸馏模型对病害区域的识别能力显著增强。该方法在精度与效率间取得了良好平衡，适合边缘设备部署，可为农业现代化提供有效技术支撑。

Abstract: Plant diseases are a major threat to global food security. Traditional manual inspection methods cannot meet the demands of modern smart agriculture for accurate and automated disease identification. Deep learning offers a promising solution, yet conventional convolutional neural networks suffer from large parameter sizes and high computational complexity, while lightweight networks often face insufficient classification accuracy. To address this issue, this paper proposes a plant disease classification method based on multi-level logit knowledge distillation. Taking ResNet34 as the teacher model, the method combines instance-level, batch-level, and class-level distillation losses to transfer knowledge from the teacher network to lightweight student models from multiple dimensions, thereby improving their classification performance without increasing inference overhead. Experiments are conducted on the publicly available multi-source Plant Disease Classification Merged Dataset from Kaggle. The results show that the distilled student model achieves a Top-1 accuracy of up to 95.19%, surpassing the teacher model’s 94.20%. CAM visualization further confirms that the distilled model’s ability to recognize disease regions is significantly enhanced. The proposed method achieves a favorable balance between accuracy and efficiency, is suitable for deployment on edge devices, and can provide effective technical support for agricultural modernization.