Table 1 Ablation study of the proposed modules, including computational efficiency metrics. “Gan” is an abbreviation for the RepNCSPELAN4 module.
From: Balancing complexity and accuracy for defect detection on filters with an improved RT-DETR
Model configuration | Params(M) | FLOPs(G) | Precision(%) | Recall(%) | mAP@0.5(%) | mAP@0.5:0.95(%) | FPS |
|---|---|---|---|---|---|---|---|
Baseline (RT-DETR) | 20.1 | 58.29 | 83.20 | 90.54 | 90.32 | 65.37 | 134.1 |
Baseline + LSKA | 20.9 | 58.29 | 97.33 | 94.72 | 95.55 | 70.59 | 119.8 |
Baseline + Gan (RepNCSPELAN4) | 18.9 | 51.94 | 94.73 | 93.22 | 93.87 | 69.54 | 171.9 |
Baseline + Adown | 17.2 | 50.89 | 86.44 | 92.29 | 91.20 | 68.21 | 134.7 |
Baseline + LSKA + Gan | 19.6 | 52.07 | 98.53 | 96.76 | 97.86 | 76.26 | 120.1 |
Final model (Baseline + LSKA + Gan + Adown) | 18.7 | 50.67 | 98.54 | 97.07 | 97.67 | 75.72 | 118.1 |
