Page 90 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

74

Through the comparison table, we see that the YOLOv7-E6 model is much better than
the models compared (YOLOv5-m, YOLOv7-w6, YOLOv8-x) in terms of mAP, Re-
call parameters but inferior to all 3 models above in terms of Precision and training time
and memory used.

Fig. 8. Detection results are obtained from YOLOv7-E6

Model Evaluation. We will analyze the results obtained for the model with the highest
result, YOLOv7- E6.

Fig. 9. PR Curve và F1 Curve

The model trained with YOLOv7-E6 showed an average mAP accuracy of 0.891, which
is 89.1% correct on the test set.
Regarding the F1 curve, the model reaches 0.86 when the confidence is set at 0.374.

CITA 2023 ISBN: 978-604-80-8083-9