Table 11.
Inference time analysis of the CNN models compared in Table 10 recorded on NVIDIA Tesla K80 GPUs. In all the compared methods, ResNet50 was used as the common backbone.
| S. No. | Method | Backbone | Inference time (milliseconds/image) | Parameters (in millions) | Giga FLOPs |
|---|---|---|---|---|---|
| 1 | U-Net [39] | ResNet50 encoder | 34.37 | 32.51 | 10.56 |
| 2 | Attention U-net [40] | ResNet50 encoder | 47.12 | 36.07 | 12.45 |
| 3 | R2U Net [41] | ResNet50 | 57.80 | 27.31 | 15.80 |
| 4 | FCN8s [42] | ResNet50 | 25.54 | 26.10 | 7.71 |
| 5 | DeCovNet [47] | ResNet50 | 20.25 | 22.12 | 6.36 |
| 6 | DeepLabV3 [43] | ResNet50 | 62.09 | 39.62 | 40.71 |
| 7 | Link Net [44] | ResNet50 | 27.25 | 31.17 | 10.63 |
| 8 | PSPNet [45] | ResNet50 | 11.05 | 24.29 | 2.83 |
| 9 | Proposed CNN | Inception-ResNet-V2 based MKE module | 38.24 | 30.51 | 13.78 |