Performance Analysis of YOLOv8–YOLOv12 for Edge-Based Strawberry Ripeness Detection

This study presents a comprehensive benchmark of the most recent YOLO object detection architectures, from YOLOv8 to YOLOv12, applied to the binary visual classification of ripe and unripe strawberries, a relevant task in precision agriculture and automated post-harvest processing. Using a controlled, high-resolution dataset of 1,364 labeled images and a standardized training pipeline, we evaluate lightweight variants of each model (nano, tiny, and small), focusing on performance indicators essential for real-time edge deployment, including precision, recall, F1-score, mAP{@}0.5,$\mathbf{m A P} {@} 0.5: 0.95$, and per-image inference time. Results show that the YOLOv9-tiny model achieves the best overall performance, with perfect recall (1.0000), an F1-score of 0.9550, and the highest$\text{mAP} {@} 0.5: 0.95(0.9613)$, outperforming both earlier and more recent architectures. YOLOv12 variants demonstrate strong generalization and high sensitivity, but with slightly lower precision, indicating a higher tendency for false positives, a critical aspect in automated sorting systems. The study also reveals that newer versions do not necessarily surpass older ones across all metrics, emphasizing that architectural complexity alone does not guarantee superior performance. The main contribution of this work is a unified and fair comparison of five generations of YOLO models under the same hardware, dataset, and hyperparameter conditions, providing new insights into how architectural innovations, such as PGI, NMS-free design, enhanced attention mechanisms, and lightweight transformer components, affect performance in agricultural classification tasks. These findings offer practical guidance for selecting lightweight models for edge-based fruit quality assessment and highlight research opportunities to integrate multispectral sensors, domain adaptation techniques, and real-world evaluations into future studies.