Automatic diagnosis of keratitis using object localization combined with cost-sensitive deep attention convolutional neural network

Journal of Big Data

Table 2 Comparison of AP and mAP performance of eight object localization methods on four datasets

Methods		NEH			JEH			NOC			ZEH
Methods		Cornea	Conj_Cor	mAP	Cornea	Conj_Cor	mAP	Cornea	Conj_Cor	mAP	Cornea	Conj_Cor	mAP
Two-stage localization methods	Faster R-CNN1	0.9894	1.0000	0.9947	0.9889	0.9986	0.9938	0.9897	1.0000	0.9949	0.9897	1.0000	0.9948
	Faster R-CNN2	0.9894	1.0000	0.9947	0.9890	0.9900	0.9895	0.9791	1.0000	0.9895	0.9897	0.9994	0.9946
	Cascade R-CNN1	0.9894	1.0000	0.9947	0.9900	0.9898	0.9899	0.9899	0.9999	0.9949	0.9901	0.9901	0.9901
	Cascade R-CNN2	0.9894	1.0000	0.9947	0.9901	0.9898	0.9899	0.9898	0.9999	0.9948	0.9901	0.9901	0.9901
	TridentNet	0.9894	1.0000	0.9947	0.9599	0.9898	0.9748	0.9801	1.0000	0.9901	0.9801	1.0000	0.9901
One-stage localization methods	RetinaNet1	0.9894	1.0000	0.9947	0.9972	0.9970	0.9971	0.9983	1.0000	0.9992	0.9997	0.9993	0.9995
	RetinaNet2	0.9896	1.0000	0.9948	0.9987	0.9957	0.9972	0.9900	0.9998	0.9949	0.9999	0.9997	0.9998
	SSD	0.9898	1.0000	0.9949	0.9901	0.9901	0.9901	0.9901	1.0000	0.9951	0.9999	1.0000	1.0000

Bold values represent the best performance among different methods in the same column
AP average precision, mAP mean average precision, Conj_Cor conjunctiva and cornea