Scholarlink

<p>Potato disease recognition has emerged as a high-impact application of deep learning in precision agriculture. However, existing datasets exhibit severe class imbalance, which adversely affects the generalization ability of standard classifiers. To address this challenge, we introduce a conditional GAN named FFRA-CGAN (Frequency-aware and Focal Residual-Attention CGAN), along with a two-stage training strategy tailored to potato pathology. In the first stage, the conditional generator, augmented with residual connections and self-attention, produces structurally faithful potato leaf images. The training process is guided by Focal Frequency Loss (FFL), which enforces frequency-domain alignment to retain vein topology and lesion morphology. In the second stage, the discriminator pretrained on a balanced mix of real and frequency-aligned synthetic images is reused as an independent classifier, thereby eliminating the need for explicit resampling or augmentation. Extensive experiments on the PlantVillage potato dataset demonstrate that FFRA-CGAN achieves state-of-the-art accuracy, significantly improving recognition of the underrepresented healthy class. This study demonstrates that integrating frequency-aligned image generation with discriminator reuse provides a robust and generalizable solution to long-tailed recognition challenges in agricultural disease detection, supporting improved decision-making in crop management.</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

Frequency-Aware and Residual-Attention CGAN with Two-Stage Training for Imbalanced Potato Leaf Disease Detection in Crop Management

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • An Zhang,
  • Chao Wu,
  • Guiyuan Li,
  • Sheng Chen

摘要

Potato disease recognition has emerged as a high-impact application of deep learning in precision agriculture. However, existing datasets exhibit severe class imbalance, which adversely affects the generalization ability of standard classifiers. To address this challenge, we introduce a conditional GAN named FFRA-CGAN (Frequency-aware and Focal Residual-Attention CGAN), along with a two-stage training strategy tailored to potato pathology. In the first stage, the conditional generator, augmented with residual connections and self-attention, produces structurally faithful potato leaf images. The training process is guided by Focal Frequency Loss (FFL), which enforces frequency-domain alignment to retain vein topology and lesion morphology. In the second stage, the discriminator pretrained on a balanced mix of real and frequency-aligned synthetic images is reused as an independent classifier, thereby eliminating the need for explicit resampling or augmentation. Extensive experiments on the PlantVillage potato dataset demonstrate that FFRA-CGAN achieves state-of-the-art accuracy, significantly improving recognition of the underrepresented healthy class. This study demonstrates that integrating frequency-aligned image generation with discriminator reuse provides a robust and generalizable solution to long-tailed recognition challenges in agricultural disease detection, supporting improved decision-making in crop management.