Desi Fatkhi Azizah, Anik Nur Handayani, Aji Prasetya Wibawa
This study evaluates the performance of four YOLOv8 variants (n, s, m, l) for detecting SIBI sign language with a mono background under two dataset scenarios: without augmentation and with augmentation. Augmentation was applied through orientation transformation, rotation, brightness adjustment, blur, and noise to increase data variation and improve model generalization. All models were trained for 50 epochs to ensure optimal convergence. Results show that augmentation generally improves performance, especially for YOLOv8s and YOLOv8l. YOLOv8s achieved notable gains in F1-Score (0.980) and Precision (0.9829), while YOLOv8l recorded the highest mAP50 (0.9908). YOLOv8n maintained the fastest inference time (4.15 ms) but with smaller accuracy improvements. A trade-off between accuracy and inference speed was observed, with larger models offering higher accuracy but slower processing. Dataset augmentation proved effective, particularly for medium to large models. Model choice should align with application needs: YOLOv8l for accuracy, YOLOv8n for speed, and YOLOv8s for balanced performance. ©2025 IEEE.
Department of Electrical Engineering and Informatics, Universitas Negeri Malang, Malang, Indonesia
