Table 2 Performance comparison in terms of mAP (%), F1, precision (%), and recall (%) between our proposed method and the state-of-the-art algorithms on CLD dataset in the 30-shot scenario.
From: Sparse cross-transformer network for surface defect detection
Method | Backbone | Metric | mAP of each class | |||||
|---|---|---|---|---|---|---|---|---|
mAP | F1 | Precision | Recall | Scratch | Sand hole | Wear | ||
MetaRCNN | ResNet101 | 52.42 | 0.6648 | 61.52 | 72.31 | 60.81 | 45.63 | 50.82 |
TFA | ResNet101 | 51.73 | 0.6501 | 72.05 | 59.22 | 58.27 | 45.89 | 51.03 |
FSCE | ResNet101 | 52.61 | 0.6829 | 62.56 | 75.18 | 58.93 | 52.16 | 46.74 |
AttentionRPN | ResNet50 | 49.86 | 0.5982 | 65.39 | 55.13 | 50.14 | 40.92 | 58.52 |
FSDetView | ResNet101 | 47.90 | 0.7211 | 65.82 | 79.74 | 52.17 | 42.83 | 48.70 |
MPSR | ResNet101 | 55.91 | 0.7054 | 61.63 | 82.46 | 52.37 | 51.53 | 63.83 |
QA-FewDet | ResNet101 | 51.64 | 0.7378 | 76.28 | 71.43 | 56.74 | 49.27 | 48.91 |
DeFRCN | ResNet101 | 58.23 | 0.7375 | 69.82 | 78.15 | 57.39 | 56.08 | 61.22 |
FCT | PvTv2 | 59.81 | 0.7889 | 75.06 | 83.13 | 60.11 | 56.62 | 62.7 |
Ours | PvTv2 | 62.73 | 0.8057 | 76.38 | 85.24 | 65.06 | 59.88 | 63.25 |
