<![CDATA[This study conducts a comparative performance evaluation of lightweight and medium variants of YOLO models, specifically YOLOv8, YOLOv10, and YOLOv12, for real-time chili ripeness detection using digital image analysis. A dataset comprising 1,450 images of Baskara chili peppers at four ripeness stages, namely green, yellow, orange, and red, was collected using a conveyor-based imaging system under controlled lighting conditions. The images were manually annotated with bounding boxes and divided into training, validation, and test sets in proportions of 73.1%, 18.3%, and 8.6%, respectively. All models were trained with identical parameters to ensure a fair comparison and evaluated using precision, recall, F1-score, mean Average Precision, and computational efficiency metrics. The results indicate that YOLOv12s achieved the highest overall performance, with a precision of 0.918, recall of 0.938, mAP@0.5 of 0.960, mAP@0.5:0.95 of 0.865, F1-score of 0.927, 21.2 GFLOPs, and an inference time of 3.1 ms. Evaluation on 125 additional images confirmed robust generalization, with a precision of 0.908, recall of 0.923, mAP@0.5 of 0.931, and mAP@0.5:0.95 of 0.849. Class-wise analysis showed that the green class achieved the highest detection accuracy, while the orange class was the most challenging due to visual similarity with adjacent ripeness stages. Overall, YOLOv12s achieved an optimal balance between detection accuracy and computational efficiency, making it promising for real-time chili sorting in smart agriculture applications.]]>
