Scholarlink

<p>We present a dual-pass deep learning framework for real-time citrus leaf disease analysis that combines a Vision Transformer–U-Net segmentation pipeline with an adaptive ensemble classifier. In the first pass, the ViT–U-Net backbone segments lesions with a mean Intersection over Union of 0.66 and Dice coefficient of 0.78—improving upon a standard U-Net baseline (IoU 0.62, Dice 0.74). A secondary color-threshold refinement sharpens lesion boundaries. We then fuse deep semantic embeddings with handcrafted color, texture, and shape descriptors, and feed them into an ensemble of CNN and transformer heads. This classifier achieves 91.30% overall accuracy—surpassing CNN-only (89.13%) and ViT-only (86.96%) counterparts—and attains a macro-averaged F1 score of 85.76%. Extensive experiments demonstrate robust performance across varied leaf shapes, lighting conditions, and disease categories, establishing a data-efficient, interpretable solution suited to resource-constrained agricultural settings.</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

A dual-pass Vision Transformer–U-Net framework with adaptive feature fusion for citrus leaf disease segmentation and classification

  • 期刊论文

  • 发表日期：

    2025年8月22日

查看全文
  • Archna Goyal,
  • Kamlesh Lakhwani

摘要

We present a dual-pass deep learning framework for real-time citrus leaf disease analysis that combines a Vision Transformer–U-Net segmentation pipeline with an adaptive ensemble classifier. In the first pass, the ViT–U-Net backbone segments lesions with a mean Intersection over Union of 0.66 and Dice coefficient of 0.78—improving upon a standard U-Net baseline (IoU 0.62, Dice 0.74). A secondary color-threshold refinement sharpens lesion boundaries. We then fuse deep semantic embeddings with handcrafted color, texture, and shape descriptors, and feed them into an ensemble of CNN and transformer heads. This classifier achieves 91.30% overall accuracy—surpassing CNN-only (89.13%) and ViT-only (86.96%) counterparts—and attains a macro-averaged F1 score of 85.76%. Extensive experiments demonstrate robust performance across varied leaf shapes, lighting conditions, and disease categories, establishing a data-efficient, interpretable solution suited to resource-constrained agricultural settings.