Scholarlink

<p>Limited disease image datasets and insufficient sample diversity significantly constrain the performance of deep learning models in agricultural plant disease identification applications. In this study, we developed UPSGAN, an enhanced CycleGAN-based framework that pioneers a novel paradigm by generating diseased samples from abundant healthy leaf images rather than augmenting existing diseased specimens.Our framework introduces an improved U-Net generator architecture integrated with Progressive Lightweight Multi-branch Group (PLMG) modules that enable dynamic weight allocation across four attention branches, significantly enhancing multi-scale feature extraction while preserving critical structural details.The discriminator incorporates strategic spectral normalization and self-attention mechanisms to achieve superior training stability and global feature modeling capabilities.Comprehensive experiments on the PlantVillage dataset demonstrate that UPSGAN achieves a state-of-the-art FID of 46.94 ± 1.58, a 33.5% improvement over the baseline CycleGAN (70.57 ± 6.90). It also shows a 58.6% improvement in KID(0.0257 ± 0.0191 vs. 0.0621 ± 0.0099) and a 33.7% increase in IS (2.05 ± 0.21 vs. 1.54 ± 0.07). Critically, in downstream classification tasks across eight diverse architectures, UPSGAN-augmented data yields consistent accuracy improvements, averaging 2.05% with a peak gain of 2.82% on DenseNet121. The model maintains practical feasibility with a moderate computational overhead (38.6% increase in FLOPs). UPSGAN establishes a viable and data-efficient solution for addressing sample scarcity in precision agriculture.</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

Enhancing tomato disease identification via UPSGAN: an advanced cycleGAN-based image augmentation approach

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Yuqing Zhou,
  • Yonghong Wu

摘要

Limited disease image datasets and insufficient sample diversity significantly constrain the performance of deep learning models in agricultural plant disease identification applications. In this study, we developed UPSGAN, an enhanced CycleGAN-based framework that pioneers a novel paradigm by generating diseased samples from abundant healthy leaf images rather than augmenting existing diseased specimens.Our framework introduces an improved U-Net generator architecture integrated with Progressive Lightweight Multi-branch Group (PLMG) modules that enable dynamic weight allocation across four attention branches, significantly enhancing multi-scale feature extraction while preserving critical structural details.The discriminator incorporates strategic spectral normalization and self-attention mechanisms to achieve superior training stability and global feature modeling capabilities.Comprehensive experiments on the PlantVillage dataset demonstrate that UPSGAN achieves a state-of-the-art FID of 46.94 ± 1.58, a 33.5% improvement over the baseline CycleGAN (70.57 ± 6.90). It also shows a 58.6% improvement in KID(0.0257 ± 0.0191 vs. 0.0621 ± 0.0099) and a 33.7% increase in IS (2.05 ± 0.21 vs. 1.54 ± 0.07). Critically, in downstream classification tasks across eight diverse architectures, UPSGAN-augmented data yields consistent accuracy improvements, averaging 2.05% with a peak gain of 2.82% on DenseNet121. The model maintains practical feasibility with a moderate computational overhead (38.6% increase in FLOPs). UPSGAN establishes a viable and data-efficient solution for addressing sample scarcity in precision agriculture.