Research Article
FWDNet: A Novel Recognition Network for Ferrography Wear Debris Image Analysis
Table 2
The contrast of experimental results of 5 DCNN models.
| | DCNN models | Fine-tuned layers | Time cost per epoch (s) | Validation accuracy (%) | Test accuracy (%) |
| | AlexNet | The whole network | 745 | 76.00 | 74.60 |
| | VGG16 | Dense net | 759 | 75.00 | | | Block 5 | 760 | 81.00 | | | Block 4 | 766 | 83.00 | | | Block 3 | 759 | 85.00 | | | Block 2 | 763 | 83.00 | | | The whole network | 759 | 84.00 | 83.99 |
| | InceptionV3 | Dense net | 770 | 70.00 | | | Block 5 | 784 | 85.00 | | | Block 4 | 791 | 88.00 | | | Block 3 | 794 | 89.00 | | | The whole network | 798 | 88.00 | 82.19 |
| | ResNet50 | Dense net | 409 | 44.00 | | | Block 5 | 420 | 89.00 | | | Block 4 | 430 | 89.00 | | | Block 3 | 437 | 91.00 | | | The whole network | 441 | 92.00 | 87.99 |
| | DenseNet121 | Dense net | 413 | 72.00 | | | Block 5 | 437 | 86.00 | | | Block 4 | 451 | 89.00 | | | Block 3 | 486 | 93.00 | | | The whole network | 535 | 92.00 | 88.39 |
|
|
