EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

A robust fusion–pruning framework for lightweight CNNs: Application to apple leaf disease diagnosis under real-world image corruptions

Oluwaseun O. Martins, Christiaan C. Oosthuizen, Dawood A. Desai

Results in Engineering · 2026-06 · 30 · 110612

DOI: 10.1016/j.rineng.2026.110612 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Results in Engineering, 2026-06. Verfasst von Oluwaseun O. Martins und 2 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 79 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Tshwane University of Technology.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Timely and reliable plant disease detection is essential for precision agriculture; however, achieving robustness under real-world image degradations while maintaining edge deployability remains challenging. This study proposes a unified fusion–pruning framework that integrates complementary feature representations from MobileNetV2 and NASNetMobile, combined with unstructured magnitude-based pruning to enhance robustness and computational efficiency. The framework is evaluated on a four-class apple leaf dataset (3,171 images) derived from the PlantVillage dataset, with cross-domain validation on the PlantDoc dataset. Baseline models achieve high clean-set accuracy (up to 97.49%) but exhibit notable degradation under corruption. In contrast, the proposed pruned fusion model attains 97.81% accuracy and consistently outperforms baseline models across common image corruptions, including noise, blur, and illumination distortions. Statistical validation using ANOVA and cross-validation confirms the significance and consistency of these improvements. Robustness metrics, including Area Under the Curve (AUC) and mean Corruption Error (mCE), further demonstrate monotonic performance gains with increasing sparsity, identifying an optimal pruning range of 50–70%, where sparsification effectively acts as an implicit feature selection mechanism. To enhance interpretability and trust, Grad-CAM++-based attention analysis is incorporated, with quantitative evaluation of attention alignment demonstrating improved localization of disease-relevant regions in the pruned fusion model. From a deployment perspective, the 70% pruned model achieves an effective trade-off between robustness and efficiency (31.49 MB), while post-training quantization reduces the model size to 8.44 MB, enabling deployment in resource-constrained environments. Overall, the proposed framework provides a scalable, interpretable, and reproducible solution for corruption-resilient edge AI in plant disease diagnosis.

Förderung

  • Tshwane University of Technology

Zitieren

APA

Martins, O. O., Oosthuizen, C. C., & Desai, D. A. (2026). A robust fusion–pruning framework for lightweight CNNs: Application to apple leaf disease diagnosis under real-world image corruptions. Results in Engineering, 30, 110612. https://doi.org/10.1016/j.rineng.2026.110612

Vancouver

Martins OO, Oosthuizen CC, Desai DA. A robust fusion–pruning framework for lightweight CNNs: Application to apple leaf disease diagnosis under real-world image corruptions. Results in Engineering. 2026;30:110612. doi:10.1016/j.rineng.2026.110612

BibTeX
@article{Martins2026,
  author = {Martins, Oluwaseun O. and Oosthuizen, Christiaan C. and Desai, Dawood A.},
  title = {A robust fusion–pruning framework for lightweight CNNs: Application to apple leaf disease diagnosis under real-world image corruptions},
  journal = {Results in Engineering},
  year = {2026},
  volume = {30},
  pages = {110612},
  doi = {10.1016/j.rineng.2026.110612}
}
Kostenlos starten →