Work

CoBiOpt-EFA: Confidence-guided Bilevel Optimization With Early-feature-aware Loss for Early-stage Rice Symptom Recognition

Year: 2026
Type: article
Abstract: CoBiOpt-EFA (Confidence-Guided Bilevel Optimization with Early-Feature-Aware Loss) is a deep learning framework for rice disease classification from visually subtle or early-observable symptoms.The framework integrates YOLOv11n-based region-of-interest localization and MobileNetV3-Small feature extraction with Early Feature Importance Estimation (EFIE), Confidence-Guided Learning (CGL), Early-Feature-Aware (EFA) loss, and bilevel optimization.Experiments were conducted using a validated dataset of 4,820 unique rice leaf images representing four disease classes, with 4,320 images used for the fixed classification experiment.Under the representative seed-999 evaluation, CoBiOpt-EFA achieved 0.984 accuracy, 0.985 macro-precision, 0.984 macrorecall, and 0.984 macro-F1, with 0.929 M trainable parameters.Across five random seeds, the mean accuracy and macro-F1 were 0.957 ± 0.017 and 0.958 ± 0.017, respectively, with 95% confidence intervals of 0.935-0.979and 0.937-0.979,indicating measurable run-to-run variability.Robustness analysis showed sensitivity to blur, illumination, contrast, and noise, while detector-confidence analysis indicated that confidence was not a direct proxy for classification correctness.The classifier required 0.061 GFLOPs and 6.137 ± 1.568 ms/image, while the complete detector-inclusive pipeline required 24.765 ± 1.877 ms/image.These results characterize the performance and computational behavior of CoBiOpt-EFA under the evaluated conditions and highlight the need for broader validation under diverse field conditions.
Language: English
Subfield: Plant Science
Open Access status: hybrid
OpenAlex ID: W7234780634
FWCI:
Cites: 39
Cited by:
Related to: 10