Table 3.
Experiments on recently published biological image classification tasks
| Data sets | Categories | Our method | Original study | Method | Reference |
|---|---|---|---|---|---|
| ClearedLeaf | 19 families | 88.7 | 71 | SIFT + SVM | (Wilf et al., 2016) |
| CRLeaves | 255 species | 94.67 | 51 | finetune InceptionV3 | (Carranza-Rojas et al., 2017) |
| EcuadorMoths | 675 species | 55.4 (58.2) |
55.7 57.19 |
AlexNet + SVM VGG16 + SCDA + SVM |
(Rodner et al., 2015) (Wei et al., 2017) |
| Flavia | 32 species | 99.95 | 99.65 | ResNet26 | (Sun et al., 2017) |
| Pollen23 | 23 species | 94.8 | 64 | CST + BOW | (Gonçalves et al., 2016) |
Note: We used input images of size 
(
for Pollen23), global average pooling, fusion of c1–c5 features, and signed square root normalization. Accuracy is reported using 10-fold cross validation except for EcuadorMoths, where we used the same partitioning as in the original study. Bold font indicates the best identification performance.