Deep Transfer Learning for Biology Cross-Domain Image Classification

Table 12

Results of cross-domain transfer learning on Plant Seedlings.

		Transfer process			Model	Accuracy (%)
		Flowers17		Plant Seedlings	GoogLeNet-v3	97.75	0.9762
		Flowers102		Plant Seedlings	GoogLeNet-v3	97.64	0.9750
		PlanktonSet 1.0		Plant Seedlings	GoogLeNet-v3	97.47	0.9724
		QUT Fish		Plant Seedlings	GoogLeNet-v3	97.32	0.9702
ImageNet		Flowers17		Plant Seedlings	GoogLeNet-v3	98.06	0.9784
ImageNet		Flowers102		Plant Seedlings	GoogLeNet-v3	98.16	0.9806
ImageNet		PlanktonSet 1.0		Plant Seedlings	GoogLeNet-v3	97.58	0.9738
ImageNet		QUT Fish		Plant Seedlings	GoogLeNet-v3	97.85	0.9765
		Flowers17		Plant Seedlings	ResNet-18	98.00	0.9775
		Flowers102		Plant Seedlings	ResNet-18	97.77	0.9750
		PlanktonSet 1.0		Plant Seedlings	ResNet-18	97.79	0.9751
		QUT Fish		Plant Seedlings	ResNet-18	97.31	0.9700
ImageNet		Flowers17		Plant Seedlings	ResNet-18	97.96	0.9777
ImageNet		Flowers102		Plant Seedlings	ResNet-18	98.11	0.9791
ImageNet		PlanktonSet 1.0		Plant Seedlings	ResNet-18	97.52	0.9720
ImageNet		QUT Fish		Plant Seedlings	ResNet-18	98.00	0.9778
		Flowers17		Plant Seedlings	ResNet-34	97.96	0.9771
		Flowers102		Plant Seedlings	ResNet-34	97.79	0.9755
		PlanktonSet 1.0		Plant Seedlings	ResNet-34	97.90	0.9767
		QUT Fish		Plant Seedlings	ResNet-34	97.31	0.9701
ImageNet		Flowers17		Plant Seedlings	ResNet-34	98.11	0.9789
ImageNet		Flowers102		Plant Seedlings	ResNet-34	98.27	0.9813
ImageNet		PlanktonSet 1.0		Plant Seedlings	ResNet-34	97.75	0.9749
ImageNet		QUT Fish		Plant Seedlings	ResNet-34	98.19	0.
		Flowers17		Plant Seedlings	ResNet-50	97.90	0.9760
		Flowers102		Plant Seedlings	ResNet-50	97.54	0.9720
		PlanktonSet 1.0		Plant Seedlings	ResNet-50	96.21	0.9581
		QUT Fish		Plant Seedlings	ResNet-50	97.24	0.9694
ImageNet		Flowers17		Plant Seedlings	ResNet-50	98.32	0.9822
ImageNet		Flowers102		Plant Seedlings	ResNet-50	98.25	0.9807
ImageNet		PlanktonSet 1.0		Plant Seedlings	ResNet-50	96.63	0.9630
ImageNet		QUT Fish		Plant Seedlings	ResNet-50	98.29	0.9814

indicates the results outperform the corresponding results of training from scratch and fine-tuning on ImageNet.