Fig. 1. CycleGAN architecture
Fig. 2. DAM-CycleGAN algorithm flowchart
Fig. 3. CycleGAN network architecture based on dual attention mechanism
Fig. 4. Generator_DA model structure based on dual attention mechanism
Fig. 5. Experimental data. (a) Original dataset; (b) dataset used in this paper; (c) Google Maps converted into sea-land binarized image by sea-land binarization operation
Fig. 6. Converted sea-land binarization effect maps (100 pairs of training set)
Fig. 7. Experimental results of different models (300 pairs of training sets)
Method | MSE | MPA | MIoU |
---|
Method one | 2648.66 | 0.8219 | 0.8956 | Method two | 2052.87 | 0.8953 | 0.9179 | Method three | 2655.68 | 0.8922 | 0.9074 | Method four | 2047.31 | 0.9100 | 0.9241 | Method five | 1642.98 | 0.9201 | 0.9363 |
|
Table 1. Comparison results of MSE, MPA, and MIoU on 35 pairs of training sets and 50 pairs of test sets
Method | MSE | MPA | MIoU |
---|
Method one | 4207.07 | 0.8208 | 0.8389 | Method two | 1857.45 | 0.9013 | 0.9157 | Method three | 1323.26 | 0.9305 | 0.9384 | Method four | 1400.98 | 0.9178 | 0.9268 | Method five | 1126.54 | 0.9258 | 0.9390 |
|
Table 2. Comparison results of MSE, MPA, and MIoU on 100 pairs of training sets and 50 pairs of test sets
Model | MSE | MPA | MIoU |
---|
FCN | 3685.87 | 0.9384 | 0.8608 | DeepLab | 3761.87 | 0.9351 | 0.8736 | DAM-CycleGAN | 1043.04 | 0.9310 | 0.9533 |
|
Table 3. Comparison results of MSE, MPA, and MIoU on 200 pairs of training sets and 30 pairs of test sets
Model | MSE | MPA | MIoU |
---|
FCN | 4533.15 | 0.9259 | 0.8852 | DeepLab | 3125.88 | 0.9453 | 0.8944 | DAM-CycleGAN | 946.40 | 0.9335 | 0.9568 |
|
Table 4. Comparison results of MSE, MPA, and MIoU on 300 pairs of training sets and 30 pairs of test sets