Modifications | Model | Precision (%) | Recall rate (%) | F1-score (%) | mAP (%) | Training time (h) | Inference time (per image) (ms) | B-FLOPS | Size (MB) |
---|---|---|---|---|---|---|---|---|---|
Original | YOLOv4 | 84 | 95 | 89 | 93.87 | 48 | 726.66 | 59.57 | 244.40 |
Residual block pruning | YOLOv4-RC3 | 84 | 92 | 88 | 91.65 | 35 | 678.53 | 47.59 | 242.40 |
YOLOv4-RC4 | 83 | 92 | 87 | 92.84 | 37 | 703.82 | 51.21 | 233.20 | |
YOLOv4-RC5 | 85 | 89 | 87 | 92.47 | 37 | 704.48 | 57.61 | 222.10 | |
YOLOv4-RC3_4 | 83 | 89 | 86 | 88.09 | 32 | 676.18 | 37.35 | 221.50 | |
YOLOv4-RC3_5 | 77 | 77 | 77 | 76.56 | 32.5 | 680.01 | 45.64 | 220.4 | |
Backbone replacement | YOLOv4- ResNet-50L | 70 | 84 | 76 | 79.70 | 28 | 719.50 | 37.33 | 209.30 |
YOLOv4-ResNet-50Â M | 74 | 86 | 80 | 81.43 | 28 | 884.82 | 37.33 | 209.30 |