Page 89 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

73

model were evaluated by the mAP, Precision, Recall, Training time and memory
measurements to be used during training and testing.
The table of evaluation results with measurements for YOLO models V5, 7, 8 is
shown in the following table:

Table 4. Training Results
Training time Memory
Model mAP50 mAP50-95 Precision Recall
(h) (MB)
YOLOv5-n 89.9 66.2 93.3 81.9 0.669 3.8
YOLOv5-m 91 71.7 93.1 82.6 0.890 42.1
YOLOv5l 91.8 72.6 92.6 84.8 1.319 92.7
YOLOv7-X 89.7 69.4 89.3 83 2.842 142.1
YOLOv7- 92.5 73 94.9 86.4 3.904 162.4
W6
YOLOv7-E6 93.7 73.1 92.2 90.2 4.830 221.4
YOLOv8-l 89.5 69 88.5 89.1 0.991 52
YOLOv8-x 90.9 71.4 95.6 82.2 1.23 136.6

From the table of results obtained after training, we see that the results obtained give
good accuracy on the test dataset with average speed and accuracy, the reliability of the
best model is the YOLOv7-E6 version with 93.7% mAP 50, 63.1% mAP 50-95 and the
best recall with 90.2%. However, the highest precision is the YOLOv8-x with 95.6%.
YOLOv8 is the lastest and most accurate model available today.

Model Comparison. This paper aims to compare the training results of the models to
come up with a proposed model.

Table 5. Compare Training Results

Training time Memory
Model mAP50 mAP50-95 Precision Recall
(h) (MB)
YOLOv5-m 91 71.7 93.1 82.6 0.890 42.1
YOLOv7-E6 93.7 73.1 92.2 90.2 4.830 221.4
improvement +2.7 +1.4 -0.9 +7.6 +3.94 +179.3
YOLOv7-W6 92.5 73 94.9 86.4 3.904 162.4
YOLOv7-E6 93.7 73.1 92.2 90.2 4.830 221.4
improvement +1.2 +0.1 -2.7 +3.8 +0.926 +59
YOLOv8-x 90.9 71.4 95.6 82.2 1.23 136.6
YOLOv7-E6 93.7 73.1 92.2 90.2 4.830 221.4
improvement +2.8 +1.7 -3.4 +8 +3.6 +84.8

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