Table 2. The result of freezed weights, fine-tuned weights, and scratch-trained weights from Resnet-101, GoogleNet, and VGG-19 models. These results are analyzed specifically for their patterns as in Fig. 2. ā€˜lā€™ beneath each DNN model name denotes the initial learning rate of the model, which drops by 10% after every 10 epochs.

DNN model Fine or scratch Final training accuracy Final validation accuracy (%) Test accuracy (%) Training time (min:sec)
ResNet-101 (l=0.001) Freezed weight of ImageNet (20 epoch) 100 92.6 93.2 4:04
Fine-tuned (20 epoch) 100 97.2 98.1 7:48
Scratch-weightless (20 epoch) 100 95.8 96.5 8:29
Scratch-weightless (50 epoch) 100 97.8 98.2 22:48
GoogleNet (l=0.001) Freezed weight of ImageNet (30 epoch) 82 76.5 78.6 2:12
Fine-tuned (30 epoch) 100 96.7 97.2 3:21
Scratch-weightless (30 epoch) 99 96.4 96.8 3:10
Scratch-weightless (60 epoch) 100 97.6 97.6 6:27
VGG-19 (l=0.0001) Freezed weight of ImageNet (30 epoch) 80 79.46 76.59 2:24
Fine-tuned (30 epoch) 100 98.21 99.12 3:55
Scratch-weightless (30 epoch) 88 83.63 88.69 3:52
Scratch-weightless (60 epoch) 99 96.43 97.82 7:46