Table 5.
Evaluation of the segmentation-based approach using FCNNs with different training data and training strategies.
| Network | Data | Balanced | Augmented | Training Strategy | pixel acc. | mean acc. | Mean IU | f.w. IU |
|---|---|---|---|---|---|---|---|---|
| FCNN | SEM | — | — | scratch | 80.84 | 33.99 | 24.25 | 71.74 |
| FCNN | SEM | — | — | fine tuning | 92.01 | 76.26 | 68.26 | 86.80 |
| FCNN | SEM | ✓ | — | fine tuning | 92.11 | 79.63 | 66.27 | 86.91 |
| FCNN | SEM | ✓ | ✓ | fine tuning | 93.92 | 76.70 | 67.84 | 88.81 |
| FCNN | LOM | ✓ | ✓ | fine tuning | 88.27 | 54.01 | 45.43 | 81.66 |
As the results show, using SEM images, fine tuned networks and augmented data, the best performance can be achieved. However, LOM has an inferior performance compared to SEM.