<![CDATA[Visual observation of the young crescent moon is challenging due to its thin and low-contrast appearance. Although YOLO-based object detectors are promising for image-based crescent localization, newer architectures do not automatically generalize well to small grayscale datasets, and prior studies rarely report robustness across repeated training runs. This study benchmarks YOLOv8n, YOLO11n, and YOLO26n for young crescent moon detection under limited-data conditions. A grayscale dataset of 697 images was resized to 640 × 640 pixels, annotated with the single class crescent_moon, and split into training, validation, and test subsets at a fixed 70:20:10 ratio. The three models were trained using the same configuration across five random seeds. Validation results were used to analyze multi-seed robustness, while the fixed 71-image test set was used for CPU-only inference evaluation. YOLO26n achieved the highest validation mAP@50-95 and fitness with the lowest variability, and also achieved the lowest CPU pipeline latency and highest throughput on the test set. These findings show that YOLO26n offers the best trade-off between accuracy and efficiency across the evaluated dataset and CPU-only inference setting. The reported throughput reflects low-frame-rate image-based inference, not real-time video performance. This study provides a reproducible benchmark protocol that combines fixed data splitting, grayscale preprocessing, data integrity checking, multi-seed robustness analysis, and CPU inference profiling.]]>
Copyrights © 2026
