Scholarlink

Early and accurate detection of plant leaf diseases is vital for crop health. This paper proposes a disease-centric grouping strategy, focusing on shared visual features of diseases among plant species to improve generalization. We propose a novel Disease-Centric Conditional GAN for Synthetic Plant Leaf Image Disease Generation and Classification Enhancement (SLED-GAN) to generate realistic diseased leaf images conditioned on disease groups. The generator uses transposed convolutions to upsample noise and labels, while the discriminator ensures image realism and label consistency. Label smoothing improves training stability. Generated images achieve low FID and SSIM-test scores, indicating high quality and diversity. Among four classifiers, MobileNet performed best with 95.83% accuracy, which improved to 97.14% after SLED-GAN-based augmentation. The results demonstrate that the proposed data enhancement approach using SLED-GAN provides an effective solution to overfitting and significantly improves the disease identification performance.

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

SLED-GAN: A Disease-Centric Conditional GAN for Synthetic Plant Leaf Disease Image Generation and Classification Enhancement

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ivy Treesa Siby,
  • Anshita Malviya,
  • Pratik Chattopadhyay

摘要

Early and accurate detection of plant leaf diseases is vital for crop health. This paper proposes a disease-centric grouping strategy, focusing on shared visual features of diseases among plant species to improve generalization. We propose a novel Disease-Centric Conditional GAN for Synthetic Plant Leaf Image Disease Generation and Classification Enhancement (SLED-GAN) to generate realistic diseased leaf images conditioned on disease groups. The generator uses transposed convolutions to upsample noise and labels, while the discriminator ensures image realism and label consistency. Label smoothing improves training stability. Generated images achieve low FID and SSIM-test scores, indicating high quality and diversity. Among four classifiers, MobileNet performed best with 95.83% accuracy, which improved to 97.14% after SLED-GAN-based augmentation. The results demonstrate that the proposed data enhancement approach using SLED-GAN provides an effective solution to overfitting and significantly improves the disease identification performance.