Deep Learning with Taxonomic Loss for Plant Identification

Table 3

Accuracies of eight state-of-the-art neural networks trained by cross-entropy and taxonomic loss on PlantCLEF 2017 dataset.

Neural network	Loss function	Accuracy (%)
		Family	Genus	Species
GoogLeNet [13]	CL	68.73	64.22	57.86
	TAX	73.29	68.64	61.04
ResNet-50 [23]	CL	76.32	72.49	66.68
	TAX	78.95	74.77	68.04
Inception-v3 [14]	CL	77.32	73.12	67.05
	TAX	79.87	76.02	68.76
Inception-ResNet-v2 [11]	CL	79.98	75.43	68.97
	TAX	82.31	78.65	71.21
MobileNet-v2 [24]	CL	71.76	67.73	61.78
	TAX	73.88	69.40	62.01
ShuffleNet-v2 [25]	CL	61.94	57.12	49.96
	TAX	66.13	60.73	53.12
DenSeNet-169 [26]	CL	76.34	72.53	66.60
	TAX	77.87	73.92	67.10
SENet-154 [27]	CL	81.84	78.63	72.53
	TAX	84.07	79.97	73.61