Edge-Accelerated Real-Time Egg Recognition Using YOLOv8

Abstract

Mechanised egg‑collection systems require vision models that are both accurate and light enough to run on embedded hardware. We built and evaluated an end‑to‑end pipeline that couples YOLOv8 object‑detection variants with Google’s Coral Edge TPU for real‑time recognition of white and brown chicken eggs. A bespoke dataset of 971 images (640 × 480 px) was captured under diverse backgrounds and lighting, annotated in YOLO format, and split 70 %/20 %/10 % for training, validation and testing. Five YOLOv8 models (n, s, m, l, x) were trained for 100 epochs with a batch size of 16. All models achieved very high accuracy (mAP50 = 0.98), but YOLOv8s produced the best F1–confidence pairing (F1 = 0.98 at 0.703 confidence), while YOLOv8n offered the lowest computational load. Converting the networks to TensorFlow‑Lite and compiling them for the Edge TPU boosted inference speed dramatically: YOLOv8n jumped from 2.4 FPS on Raspberry Pi 5 (PyTorch) to 13.8 FPS on Edge TPU, and YOLOv8s rose from 1.0 FPS to 4.1 FPS, with only marginal accuracy loss. Precision and recall remained ≥ 0.96 across all variants. These results demonstrate that lightweight YOLOv8 models, particularly the n‑variant, are suitable for embedded, robotics‑grade egg‑collection systems that demand real‑time performance without sacrificing detection quality. Future work will expand the class set to include damaged eggs and integrate the detector into a closed‑loop robotic gripper to enable fully autonomous on‑farm operation.