基於動物辨識的YOLOv5與YOLOv8之性能比較

Performance comparison between YOLOv5 and YOLOv8 based on animal recognition

許瀟傑(Hsiao-Chieh Hsu)；吳亞芬(Ya-Fen Wu)；孫子恩(Tzu-En Sun)；林義楠(Yi-Nan Lin)

YOLOv5 ； YOLOv8 ； 深度學習 ； 物件偵測 ； 派翠網路 ； YOLOv5 ； YOLOv8 ； deep learning ； object detection ； Petri Net

明新學報

47卷（2024 / 01 / 01）

1 - 24

繁體中文;英文

本文探索比較分析YOLOv5與YOLOv8的性能及模型結構。研究中使用Anaconda3的PowerShell建立虛擬環境，運行YOLOv5和YOLOv8的程式碼，採用相同的深度模型大小（YOLOv5s、YOLOv8s）、數據集（分為兩類：dog及cat），並以固定的預測集進行比較分析。研究結果證實YOLOv5在整體準確率達到79.7%。相比之下，YOLOv8在整體準確率達到82.8%。總結而言，本文研究發現YOLOv8在準確率上優於YOLOv5，且其運行速度也遠優於後者。最後系統模擬驗證，採用一種基於數學理論和圖形化特性的系統建模工具Petri Net，分別對YOLOv5和YOLOv8的模型結構進行建模與驗證，並深入了解其兩者的架構完整性。透過分析比較證實，不同版本YOLO模型之間的性能差異，探討了有關準確性和速度之間的取捨，以及YOLOv8或未來版本的性能與改進方向。

This study explores and compares the performance and model structures of YOLOv5 and YOLOv8. Using Anaconda3's PowerShell, a virtual environment is established to run the code for YOLOv5 and YOLOv8. The study employs models of the same depth (YOLOv5s, YOLOv8s) and datasets (divided into two categories: dogs and cats), and performs comparative analysis using a fixed prediction set. The research results confirm that YOLOv5 achieves an overall accuracy of 79.7%, while YOLOv8 achieves an overall accuracy of 82.8%. In summary, this study finds that YOLOv8 outperforms YOLOv5 in accuracy and operates significantly faster. Finally, system simulation validation is conducted using a system modeling tool based on mathematical theory and graphical characteristics, Petri Net. The modeling and validation of the model structures of YOLOv5 and YOLOv8 are carried out to gain insight into their architectural integrity. Through analysis and comparison, this study confirms performance differences between different versions of YOLO models, discusses the trade-offs between accuracy and speed, and explores the performance and improvement directions of YOLOv8 or future versions.