A Metrologically Guided YOLOv12 Framework with Augmentation-Free Training and Two-Phase Optimization for Multiclass Tomato Leaf Disease Detection

Tomatoes are highly vulnerable to a wide range of leaf diseases, which significantly reduce agricultural yield and quality. Timely and precise detection of these diseases is essential for sustainable crop management and food security. This study analyzes configuration-level bidirectional multi-scale feature propagation within the native YOLOv12-s architecture, with emphasis on architectural behavior under controlled experimental conditions. The computational topology and parameterization of YOLOv12 are preserved, while bidirectional feature aggregation is activated at configuration level to examine its influence on cross-scale semantic consistency and localization reliability. The framework was trained and evaluated on a curated dataset of 4030 annotated RGB images spanning ten tomato leaf disease categories. All models were trained under an augmentation-free protocol and unified evaluation settings to isolate architectural effects from data-driven performance inflation. Under these controlled conditions, configuration-level bidirectional activation yields measurable improvements in detection consistency and spatial agreement while maintaining identical model complexity. Performance is evaluated using mAP, precision, recall, F1-score, and error-type decomposition within a measurement-consistency framework. The proposed configuration achieves 95.9% mAP@50 and 87.1% mAP@50–95 under identical experimental conditions, providing empirical evidence that topology-preserving feature routing influences multi-scale semantic stability in lesion detection.