Answers are grounded in this paper and cite their sources.
Accurate strawberry disease classification remains challenging due to visual similarities among disease categories, variations in symptom appearance, and class imbalance. Conventional whole-image classifiers may overlook localized disease-associated information that could complement global visual representations. This study proposes StrawNet-ROI, an annotation-guided dual-view deep learning framework integrating a multi-scale EfficientNet-B0 global classifier, a polygon-supervised region-of-interest (ROI) prediction network, and a separately fine-tuned local classifier. A connected-component-based procedure generates spatially constrained crops from predicted ROI maps, while confidence-adaptive probability fusion combines global and local predictions without requiring ground-truth annotations during inference. Experiments were conducted on an audited seven-class strawberry disease dataset comprising 2,457 images, partitioned into 1,426 training, 294 validation, and 737 testing samples. Across five random seeds, StrawNet-ROI achieved a mean test accuracy of 98.78 ± 0.25%, macro-precision of 98.26 ± 0.60%, macro-recall of 97.40 ± 0.43%, and macro-F1 of 97.79 ± 0.47%. Compared with the global-only classifier, the framework achieved a numerical macro-F1 improvement of 0.59 percentage points; however, the paired difference was not statistically significant (p = 0.1268). Additional cropping controls did not establish statistically significant superiority of ROI-guided selection. The complete framework contains 12.58 million parameters and achieved a mean end-to-end inference latency of 18.80 ms on a server-class GPU. These findings suggest that global–local representation fusion may provide complementary information for strawberry disease classification, although the additional computational cost and limited statistical evidence warrant caution. Because the evaluation relied on a historically examined internal test partition, independent field generalization and resource-constrained deployment remain unverified.
No takes yet. Share an insight, caveat, or question.
Al-A'arej et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: