<![CDATA[This study presents a comparative analysis of multiple YOLO generations, namely YOLOv5, YOLOv8, YOLOv11, and YOLO26, for Indonesian banknote detection under brightness variation, background clutter, and occlusion. Previous studies have mainly evaluated one or two YOLO versions under controlled conditions, while comprehensive comparisons involving newer generations such as YOLOv11 and YOLO26 for Indonesian banknote detection under real-world conditions remain limited. A dataset comprising 1,400 images of seven Indonesian banknote denominations was used for training, while 200 separately collected and manually annotated real-world images were used for evaluation. All models were trained using identical configurations with their respective pretrained weights, including 150 epochs, 640 × 640 image size, and batch size 16. Performance was evaluated using accuracy, precision, recall, F1-score, mAP@0.5, inference time, and frames per second (FPS). The results show that YOLOv11 achieves the most balanced performance, with the highest accuracy (0.950), recall (0.700), and F1-score (0.775). The mAP@0.5 values were 0.995 for YOLOv5 and 0.995 for YOLOv8, YOLOv11, and YOLO26, while the mAP@0.5:0.95 values were 0.954, 0.985, 0.987, and 0.987 for YOLOv5, YOLOv8, YOLOv11, and YOLO26, respectively. YOLO26 achieves the highest precision (0.893), while YOLOv5 provides the fastest inference time (92.463 ms) and highest FPS (10.815). Overall, YOLOv11 provides the most balanced detection performance for real-world conditions, while the results demonstrate a trade-off between detection performance and computational efficiency across YOLO generations.]]>
