Table 6 Test accuracy results using different model additions within EPNet.
From: Automated human cell classification in sparse datasets using few-shot learning
Model addition | Mini-ImageNet | BCCD (%) | HEp-2 |
|---|---|---|---|
Model addition performance | |||
No additions | 88.7% | 47.4 | 55.12% |
RandAugment (Magnitude = 5–15) | 75.8% | 34.6 | 42.1% |
RandAugment (Magnitude = 5–10) | 69.3% | 27.8 | 35.6% |
RandAugment (Magnitude = 5) | 70.1% | 28.9 | 36.1% |
Squeeze and excitation (Reduction = 0.10) | 68.7% | 27.6 | 35.2 |
Squeeze and excitation (Reduction = 0.25) | 68.2% | 26.2 | 34.9% |
Squeeze and excitation (Reduction = 2) | 63.9% | 22.4 | 30.5% |
Mixup (Alpha = 0.10) | 76.2% | 34.7 | 42.8% |
Label smoothing (A = 0.10) | 65.3 | 23.2 | 31.4% |
Exponential moving average | 78.6% | 38.2 | 44.1% |
