Table 7.

The top-1 record concerning the accuracy for the pre-trained models for NewHandPD

#	MobileNet	MobileNetV2	MobileNetV3Small	MobileNetV3Large	ResNet50	VGG16	VGG19	InceptionResNetV2
Loss function	Categorical crossentropy	Categorical crossentropy	KL divergence	KL divergence	KL divergence	KL divergence	Poisson	Categorical crossentropy
Batch size	24	8	8	24	56	40	48	56
Dropout ratio	0.42	0	0.37	0.20	0.41	0.26	0.33	0.35
TL learn ratio	89%	0%	84%	55%	22%	57	67	58
Weights optimizer	SGD Nesterov	Adam	Adagrad	SGD Nesterov	Adagrad	Adagrad	SGD	SGD Nesterov
Rotation range	${29}^{\circ }$	$0^{\circ }$	${39}^{\circ }$	$7^{\circ }$	$4^{\circ }$	${36}^{\circ }$	${31}^{\circ }$	${32}^{\circ }$
Width shift range	0.24	0	0.21	0.07	0.20	0.22	0.15	0.22
Height shift range	0.24	0	0.05	0.02	0.05	0.03	0.2	0.06
Shear range	0.02	0	0.21	0.23	0.19	0.15	0.15	0.13
Zoom range	0.12	0	0.22	0.25	0.01	0.19	0.22	0.20
Horizontal flip	$✓$	$✓$	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$
Vertical flip	$\times$	$✓$	$\times$	$\times$	$✓$	$✓$	$\times$	$\times$
Brightness range (low)	1.24	0.5	0.92	1.29	1.01	1.34	1.32	0.56
Brightness range (high)	1.28	0.5	1.08	1.52	1.67	1.76	1.46	1.3
Loss	0.038	0.032	0.152	0.107	0.049	0.029	0.180	0.049
Accuracy	99.05%	99.40%	95.12%	95.00%	98.81%	99.29%	99.75%	98.21%
F1-Score	99.05%	99.40%	95.16%	95.00%	98.81%	99.29%	99.75%	98.21%
Precision	99.05%	99.40%	95.36%	95.00%	98.81%	99.29%	99.75%	98.21%
Recall	99.05%	99.40%	95.00%	95.00%	98.81%	99.29%	99.75%	98.21%
Specificity	99.81%	99.88%	99.07%	99.00%	99.76%	99.86%	99.95%	99.64%
AUC	99.98%	99.92%	99.67%	99.82%	99.97%	99.99%	100%	99.91%
IOU coefficient	98.87%	98.72%	93.13%	94.80%	97.09%	98.02%	93.45%	97.31%
Dice coefficient	99.04%	98.99%	94.36%	95.77%	97.76%	98.50%	95.17%	97.92%