Table 2 Comparison of generalization performance in data scarcity scenarios.
Assessment indicators | MSDA-PSA | Standard U-Net | Data augmentation U-Net | Few-shot | Test condition |
|---|---|---|---|---|---|
10% data small sample IoU (%) | 72.4 | 58.6 | 65.2 | 68.7 | 5-fold cross-validation |
50% data IoU(%) | 85.3 | 76.8 | 80.1 | 78.5 | 100 epoch training |
100% data IoU (%) | 89.7 | 86.2 | 87.5 | 84.3 | Complete training set |
Cross-category generalization loU (%) | 71.2 | 53.8 | 61.5 | 64.3 | No pattern test set was seen |
Feature portability (mAP) | 0.82 | 0.68 | 0.75 | 0.79 | Downstream tasks are fine-tuned |
Training convergence speed (epoch) | 45 | 80 | 65 | 55 | LoU required 85% of the rounds |
10% data training stability (σ) | 1.8 | 3.2 | 2.6 | 2.1 | loU standard deviation |
