Research Article

Deep Transfer Learning for Biology Cross-Domain Image Classification

Table 8

Transfer learning results gains (average on GoogLeNet-v3, ResNet-18, ResNet-34 and ResNet-50) compared with the results of training from scratch.

Transfer processGain of accuracy (%)Gain of

ImageNetFlowers179.420.0911
Flowers102Flowers170.660.0070
Plant SeedlingsFlowers171.690.0150
PlanktonFlowers174.340.0423
QUT FishFlowers170.660.0053
ImageNetFlowers102Flowers179.410.
ImageNetPlant SeedlingsFlowers177.350.0727
ImageNetPlanktonFlowers176.540.0647
ImageNetQUT FishFlowers177.280.0720
ImageNetFlowers10234.830.3642
Flowers17Flowers1028.440.0871
Plant SeedlingsFlowers1023.570.0367
PlanktonFlowers10214.840.1489
QUT FishFlowers1026.310.0591
ImageNetFlowers17Flowers10233.970.3541
ImageNetPlant SeedlingsFlowers10227.620.2870
ImageNetPlanktonFlowers10221.840.2230
ImageNetQUT FishFlowers10227.670.2851
ImageNetPlant Seedlings0.530.0053
Flowers17Plant Seedlings0.210.0017
Flowers102Plant Seedlings−0.01−0.0007
PlanktonPlant Seedlings−0.35−0.0044
QUT FishPlant Seedlings−0.40−0.0051
ImageNetFlowers17Plant Seedlings0.420.0043
ImageNetFlowers102Plant Seedlings0.510.
ImageNetPlanktonPlant Seedlings−0.32−0.0041
ImageNetQUT FishPlant Seedlings0.390.0041
ImageNetPlankton−0.17−0.0005
Flowers17Plankton−0.06−0.0006
Flowers102Plankton−0.09−0.0023
Plant SeedlingsPlankton0. 0.
QUT FishPlankton−1.18−0.0189
ImageNetFlowers17Plankton−0.76−0.131
ImageNetFlowers102Plankton−0.020.
ImageNetPlant SeedlingsPlankton−0.080.
ImageNetQUT FishPlankton−0.18−0.0005
ImageNetQUT Fish20.030.1759
Flowers17QUT Fish5.380.0453
Flowers102QUT Fish1.820.0139
Plant SeedlingsQUT Fish4.300.0333
PlanktonQUT Fish11.920.1058
ImageNetFlowers17QUT Fish19.950.1725
ImageNetFlowers102QUT Fish19.720.1735
ImageNetPlant SeedlingsQUT Fish16.710.1465
ImageNetPlanktonQUT Fish15.200.1395

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