EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Severity-aware multi-stage fusion framework for leaf disease segmentation, classification, and captioning

Gunjan Thakur, Vandana Sood, Bhupesh Kumar Dewangan, Amitava Choudhury

Results in Engineering · 2026-12 · 32 · 112456

DOI: 10.1016/j.rineng.2026.112456 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Results in Engineering, 2026-12. Verfasst von Gunjan Thakur und 3 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Potato (Solanum tuberosum) cultivation is severely threatened by Early Blight and Late Blight, causing significant yield losses when not detected promptly. Although deep learning models have improved disease classification, most fail to quantify lesion severity or generate interpretable descriptions useful for farmers and agronomists. This study proposes a novel severity-aware pipeline for leaf disease detection and captioning, in which an Attention U-Net and a Severity-Aware Fusion Classifier are jointly optimized end-to-end, followed by a downstream disease-aware captioning stage conditioned on their outputs. An Attention U-Net segments diseased lesions, trained to approximate HSV-derived pseudo-masks, while suppressing background noise. A key contribution is the Severity-Aware Fusion Classifier (SAFC), which combines deep features from RGB images and lesion masks with four geometric severity descriptors: lesion area ratio, perimeter, and centroid coordinates. A Disease-Aware Captioning module further conditions a ViT-GPT2 model on predicted disease class and severity grade to generate structured, agronomically meaningful descriptions. The framework is evaluated on a balanced augmented dataset of 4,800 images covering three disease categories. An accuracy of 98.24% is achieved by the proposed SAFC, offering a favorable comparison with published PlantVillage findings. The captioning module consistently surpasses CNN-LSTM, generic ViT-GPT2, and SCOLD-style baselines across BLEU, METEOR, ROUGE-L, and CIDEr metrices. The proposed pipeline represents a promising step towards interpretable and automated potato disease diagnosis, with potential applicability in precision agriculture settings.

Zitieren

APA

Thakur, G., Sood, V., Dewangan, B. K., & Choudhury, A. (2026). Severity-aware multi-stage fusion framework for leaf disease segmentation, classification, and captioning. Results in Engineering, 32, 112456. https://doi.org/10.1016/j.rineng.2026.112456

Vancouver

Thakur G, Sood V, Dewangan BK, Choudhury A. Severity-aware multi-stage fusion framework for leaf disease segmentation, classification, and captioning. Results in Engineering. 2026;32:112456. doi:10.1016/j.rineng.2026.112456

BibTeX
@article{Thakur2026,
  author = {Thakur, Gunjan and Sood, Vandana and Dewangan, Bhupesh Kumar and Choudhury, Amitava},
  title = {Severity-aware multi-stage fusion framework for leaf disease segmentation, classification, and captioning},
  journal = {Results in Engineering},
  year = {2026},
  volume = {32},
  pages = {112456},
  doi = {10.1016/j.rineng.2026.112456}
}
Kostenlos starten →