Scholarlink

Deep learning-based plant disease detection has demonstrated remarkable accuracy, but deploying high-capacity models such as ResNet-50 in resource-constrained farm environments remains a challenge. In this work, we propose a knowledge distillation system to transfer the diagnostic capability of a highly accurate ResNet-50 teacher model to a lightweight student model, MobileNetV2, enabling effective real-world deployment. The teacher is trained on the PlantVillage dataset, and its knowledge is transferred to the student using Kullback–Leibler divergence loss with temperature scaling, alongside standard cross-entropy loss. Interestingly, the distilled student model not only replicates but outperforms the teacher, achieving a test accuracy of 99.55% compared to ResNet-50’s 99.30%, while being 7.5 \(\times \) smaller (3.4M vs. 25.5M parameters) and 4 \(\times \) faster in inference. Experimental analysis reveals that the student model corrects some of the teacher’s mistakes, suggesting that knowledge distillation acts as a regularization technique. Our findings indicate that KD can produce compact yet superior models for plant disease classification.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Distilling Deep Learning for the Field: Lightweight Plant Disease Diagnosis via ResNet-50 to MobileNetV2 Transfer

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Vishal Sharma,
  • Raju Pal

摘要

Deep learning-based plant disease detection has demonstrated remarkable accuracy, but deploying high-capacity models such as ResNet-50 in resource-constrained farm environments remains a challenge. In this work, we propose a knowledge distillation system to transfer the diagnostic capability of a highly accurate ResNet-50 teacher model to a lightweight student model, MobileNetV2, enabling effective real-world deployment. The teacher is trained on the PlantVillage dataset, and its knowledge is transferred to the student using Kullback–Leibler divergence loss with temperature scaling, alongside standard cross-entropy loss. Interestingly, the distilled student model not only replicates but outperforms the teacher, achieving a test accuracy of 99.55% compared to ResNet-50’s 99.30%, while being 7.5 \(\times \) smaller (3.4M vs. 25.5M parameters) and 4 \(\times \) faster in inference. Experimental analysis reveals that the student model corrects some of the teacher’s mistakes, suggesting that knowledge distillation acts as a regularization technique. Our findings indicate that KD can produce compact yet superior models for plant disease classification.