Karisma Trinanda Putra, Hasan Zidni, Moh. Khairudin, Adi Izhar Bin Che Ani, Rizki Ahmad Setiadi, Deny Budi Hertanto
Accurate object detection is crucial in modern poultry farming particularly for egg detection, to support automation and monitoring systems. However, deploying advanced models such as YOLOv8 on edge devices poses challenges due to limited computational and memory resources. This study evaluates the effectiveness of compressing YOLOv8n using TensorFlow Lite (TFLite) to enable real-time egg detection in resource-constrained environments. A custom dataset of annotated egg images was used to train a baseline YOLOv8n model. The trained model was then converted into two TFLite variants: 16-bit floating point (FP16) and 8-bit integer (INT8), using TensorFlow's export pipeline. The original YOLOv8n achieved 0.87 mAP@0.5, 0.88 precision, and 0.86 recall with a memory size of 5.96MB. The FP16 model preserved comparable accuracy (mAP@0.5 = 0.83) while reducing memory usage to 5.88MB and improving inference time. The INT8 model further reduced memory usage to 3.13MB, with minor accuracy degradation (mAP@0.5 = 0.80, precision = 0.85, recall = 0.84). The results demonstrate that TFLite-based compression enhances model deployability on edge hardware by reducing resource demands while maintaining real-time performance, providing practical insights for implementing efficient object detection in smart farming applications. © 2025 IEEE.
Universitas Muhammadiyah Yogyakarta, Department of Electrical Engineering, Indonesia; Education Universitas Negeri Yogyakarta Reseach, Universitas Muhammadiyah Yogyakarta, Departement of Electrical Engineering, Yogyakarta, Indonesia; Universitas Negeri Yogyakarta, Departement Of Electrical Engineering Education, Yogyakarta, Indonesia; University Teknologi MARA, Centre of Electrical Engineering Studies, Penang, Malaysia
