Table 3 Performance comparison in terms of mAP (%), F1, precision (%), and recall (%) between our proposed method and the state-of-the-art algorithms on NEU steel surface defect dataset in the 30-shot scenario.

Method	Backbone	Metric				mAP of each class
		mAP	F1	Precision	Recall	Cr	In	Pa	Ps	Rs	Sc
MetaRCNN	ResNet101	66.17	0.7086	70.64	71.09	61.53	61.53	74.19	73.24	60.17	62.8
TFA	ResNet101	65.67	0.715	76.25	67.31	62.47	76.83	71.21	63.17	58.79	61.55
FSCE	ResNet101	72.08	0.7289	65.44	82.26	63.39	75.13	68.57	78.38	76.08	70.93
AttentionRPN	ResNet50	73.81	0.6776	72.53	63.58	52.73	79.28	75.06	74.41	71.68	89.7
FSDetView	ResNet101	75.97	0.7539	71.04	80.31	73.18	82.59	63.95	78.67	84.2	73.23
MPSR	ResNet101	78.29	0.7654	68.22	87.18	75.31	80.29	65.89	86.19	79.16	82.9
QA-FewDet	ResNet101	74.01	0.7684	75.21	78.54	70.14	79.59	65.77	79.97	68.52	80.07
DeFRCN	ResNet101	80.46	0.7824	76.25	80.33	76.38	82.57	88.72	81.63	89.46	80.92
FCT	PvTv2	83.58	0.8381	80.71	87.16	81.41	82.05	83.73	82.5	89.74	76.17
Ours	PvTv2	85.29	0.8788	85.62	90.27	86.52	89.47	90.9	88.95	90.28	86.08

Quick links

Search