Scholarlink

<p>Durian leaf disease recognition requires large and diverse orchard datasets, yet such resources are scarce. Recent vision-language models (VLMs) provide new opportunities for data augmentation in precision agriculture. This study introduces DurioFLUX, a generative framework based on parameter-efficient fine-tuning (PEFT) of the FLUX model. Specifically, Low-Rank Adaptation (LoRA) is applied and further enhanced by Textual Inversion (TI) and Prompt Adapter (PA) for precise prompt alignment. Three experiments were conducted. First, a comparative study with LoRA-tuned FLUX, SD3.5, and SDXL demonstrated the superior fidelity and semantic consistency of FLUX, supported by quantitative (KID, CAS) and qualitative analyses (t-SNE, Grad-CAM). Second, in a downstream cross-domain classifier trained with real plus synthetic images, fivefold accuracy increases by + 0.15, + 0.31, and + 0.39 absolute for Algal Spot, Leaf Blight, and Leaf Spot, respectively, while No Disease class slightly (− 0.05), likely due to reflectance cues dominating in synthetic healthy leaves. Third, an ablation study confirmed complementary contributions of TI and PA beyond LoRA, particularly for minority lesions. These results demonstrate that PEFT-based generative augmentation on FLUX foundation model effectively mitigates data scarcity, enhances fine-grained disease recognition, and advances computer vision methods for agricultural classification. The code and data are available at <a href="https://github.com/wenjuanLiu518/Durian_Leaf_Disease">https://github.com/wenjuanLiu518/Durian_Leaf_Disease</a>.</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

Image augmentation by a vision-language foundation model for durian leaf disease recognition

  • 期刊论文

  • 发表日期：

    2025年10月27日

查看全文
  • Wenjuan Liu,
  • Ahmad Sufril Azlan Mohamed,
  • Mohd Azam Osman,
  • Kim Hwa Kie,
  • Chow Jeng Wong

摘要

Durian leaf disease recognition requires large and diverse orchard datasets, yet such resources are scarce. Recent vision-language models (VLMs) provide new opportunities for data augmentation in precision agriculture. This study introduces DurioFLUX, a generative framework based on parameter-efficient fine-tuning (PEFT) of the FLUX model. Specifically, Low-Rank Adaptation (LoRA) is applied and further enhanced by Textual Inversion (TI) and Prompt Adapter (PA) for precise prompt alignment. Three experiments were conducted. First, a comparative study with LoRA-tuned FLUX, SD3.5, and SDXL demonstrated the superior fidelity and semantic consistency of FLUX, supported by quantitative (KID, CAS) and qualitative analyses (t-SNE, Grad-CAM). Second, in a downstream cross-domain classifier trained with real plus synthetic images, fivefold accuracy increases by + 0.15, + 0.31, and + 0.39 absolute for Algal Spot, Leaf Blight, and Leaf Spot, respectively, while No Disease class slightly (− 0.05), likely due to reflectance cues dominating in synthetic healthy leaves. Third, an ablation study confirmed complementary contributions of TI and PA beyond LoRA, particularly for minority lesions. These results demonstrate that PEFT-based generative augmentation on FLUX foundation model effectively mitigates data scarcity, enhances fine-grained disease recognition, and advances computer vision methods for agricultural classification. The code and data are available at https://github.com/wenjuanLiu518/Durian_Leaf_Disease.