Scholarlink

<p>The pervasive threat posed by plant diseases calls for diagnostic solutions that can adapt swiftly to novel classes, even under limited data conditions. While deep learning has proven valuable in automating disease detection, its data-hungry nature can impede performance in resource-constrained settings. In this paper, we present a novel integration of the Data-Efficient Image Transformer (DeiT) and Model-Agnostic Meta-Learning++ (MAML++) for few-shot plant disease classification. By leveraging DeiT’s self-attention architecture, our framework extracts robust visual features from minimal samples, while MAML++ equips the model with the capacity for rapid adaptation to new disease categories. Experiments on a curated subset of the PlantVillage dataset demonstrate notable gains in low-data regimes: for instance, the proposed approach attains 98.8% accuracy in 2-way 5-shot classification and 97.53% accuracy in 4-way 5-shot classification of apple diseases. These results underscore the effectiveness of integrating transformers with meta-learning to handle data scarcity, offering a promising direction for precision agriculture. Key innovations include applying DeiT’s efficient feature extraction in tandem with the stability enhancements of MAML++ and validating the synergy through extensive few-shot evaluations.</p>

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

Empowering few-shot plant disease diagnosis with DeiT and MAML++

  • 期刊论文

  • 发表日期：

    2025年11月11日

查看全文
  • Metehan Günde,
  • Fesih Keskin,
  • Gültekin Işık

摘要

The pervasive threat posed by plant diseases calls for diagnostic solutions that can adapt swiftly to novel classes, even under limited data conditions. While deep learning has proven valuable in automating disease detection, its data-hungry nature can impede performance in resource-constrained settings. In this paper, we present a novel integration of the Data-Efficient Image Transformer (DeiT) and Model-Agnostic Meta-Learning++ (MAML++) for few-shot plant disease classification. By leveraging DeiT’s self-attention architecture, our framework extracts robust visual features from minimal samples, while MAML++ equips the model with the capacity for rapid adaptation to new disease categories. Experiments on a curated subset of the PlantVillage dataset demonstrate notable gains in low-data regimes: for instance, the proposed approach attains 98.8% accuracy in 2-way 5-shot classification and 97.53% accuracy in 4-way 5-shot classification of apple diseases. These results underscore the effectiveness of integrating transformers with meta-learning to handle data scarcity, offering a promising direction for precision agriculture. Key innovations include applying DeiT’s efficient feature extraction in tandem with the stability enhancements of MAML++ and validating the synergy through extensive few-shot evaluations.