Deep Learning with Taxonomic Loss for Plant Identification

Table 4

Typical images in PlantCLEF 2017 testing set and predictions made by ResNet-50 trained by two different loss functions: cross-entropy loss (CL) and taxonomic loss (TAX).

		Loss function	Family	Genus	Species
a		GT	Lilliaceae	Erythronium	Erythronium americanum Ker Gawl.
		CL	Orchidaceae	Orchis	Orchis mascula (L.) L.
		TAX	Lilliaceae	Clintonia	Clintonia andrewsiana Torr.
b		GT	Ulmaceae	Ulmus	Ulmus americana L.
		CL	Fagaceae	Fagus	Fagus grandifolia Ehrh.
		TAX	Ulmaceae	Ulmus	Ulmus crassifolia Nutt.
c		GT	Grossulariaceae	Ribes	Ribes indecorum Eastw.
		CL	Rosaceae	Holodiscus	Holodiscus discolor (pursh) Maxim.
		TAX	Grossulariaceae	Ribes	Ribes indecorum Eastw.
d		GT	Asteraceae	Heterotheca	Heterotheca canescens (DC.) Shinners
		CL	Leguminosae	Syrmatium	Syrmatium glabrum Vogel.
		TAX	Asteraceae	Heterotheca	Heterotheca canescens (DC.) Shinners

Bold values indicate the ground truth (GT) and the correct predictions.