Table 2. Comparative performance and efficiency of different artificial intelligence models for feature extractiona.
| Model | Parameter count (M) | FLOPSb (giga) | Validation accuracy (%) | mAPc | F1-score | Training time, minutes | Inference time, seconds |
|---|---|---|---|---|---|---|---|
| AlexNet | 61.1 | 0.71 d | 80.5 (79.4-81.6) |
82.8 (81.7-83.9) |
81.4 (80.3-82.5) |
151 | 39 |
| VGG16 | 138.36 | 15.47 | 83.4 (82.3-84.5) |
86.6 (85.6-87.6) |
84.5 (83.5-85.5) |
258 | 47 |
| VGG19 | 143.67 | 19.63 | 86.1 (85.1-87.1) |
86.1 (85.1-87.1) |
85.4 (84.4-86.4) |
278 | 45 |
| ResNet50 | 25.56 | 4.13 | 83.3 (82.2-84.4) |
87.2 (86.2-88.2) |
84.6 (83.6-85.6) |
258 | 59 |
| InceptionV4 | 42.62 | 6.12 | 50.3 (48.9-51.7) |
76.7 (75.5-77.9) |
66.4 (65.0-67.8) |
310 | 99 |
| DinoV2-Base | 85.51 | 21.96 | 73.3 (72.0-74.6) |
76.3 (75.1-77.5) |
73.8 (72.5-75.1) |
502 | 60 |
| Swin-Base | 87.7 | 15.17 | 86.1 (85.1-87.1) |
86.8 (85.8-87.8) |
86.4 (85.4-87.4) |
494 | 112 |
| DinoV2-Small | 21.52 | 5.52 | 70.4 (69.1-71.7) |
73.9 (72.6-75.2) |
71.1 (69.8-72.4) |
182 | 59 |
| Swin-Tiny | 28.27 | 4.37 | 60.3 (58.9-61.7) |
76.7 (75.5-77.9) |
58.8 (57.4-60.2) |
175 | 35 |
| GoogLeNet | 6.8 d | 1.51 |
95.9 (95.3-96.5)d |
89.2 (88.3-90.1)d |
95.1 (94.5-95.7)d |
57 d | 22 d |
a
Performance metrics are reported as mean (95% CI).
b
FLOPs: Floating Point Operations.
c
mAP: mean average precision.
d
The values in italics indicate the best performance or efficiency in the comparative results.