Table 8 Performance of the deployed INT8-G+MF (P+QAT) model under different class imbalance ratios. Precision (P), Recall (R), and F1-score (F1) are reported for the crack class.
From: TinyML pipeline for efficient crack classification in UAV-based structural health inspections
Class ratio (Crack:Non-crack) | Precision | Recall | F1-score |
|---|---|---|---|
1:1 (Balanced) | \(0.931 \pm 0.024\) | \(0.945 \pm 0.021\) | \(0.938 \pm 0.022\) |
1:5 | \(0.956 \pm 0.020\) | \(0.901 \pm 0.028\) | \(0.928 \pm 0.023\) |
1:10 | \(0.972 \pm 0.018\) | \(0.864 \pm 0.031\) | \(0.915 \pm 0.026\) |
