Author Affiliations
1 School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China2 School of Computer and Communications Engineering, University of Science and Technology Beijing, Beijing 100083, Chinashow less
Fig. 1. System diagram of remote sensing image classification based on DeepLab-v3+
Fig. 2. Diagrams of abtrous convolution[11]. (a) Sparse feature extraction; (b) dense feature extraction
Fig. 3. Structure of atrous spatial pyramid pooling
Fig. 4. Structure of encoder-decoder module
Fig. 5. Example of dataset. (a) Original images; (b) labels
Fig. 6. Loss curve
Fig. 7. 64 feature maps obtained from the first convolution layer
Fig. 8. Segmentation effect by DeepLab-v3+; (a) Original image; (b) segmentation result by DeepLab-v3+
Fig. 9. Comparison of loss of different models
Fig. 10. Comparison of segmentation results of DeepLab-v3+ and other models. (a) Original images; (b) segmentation results of FCN; (c) segmentation results of U-Net; (d) segmentation results of DeepLab-v3+
Fig. 11. Segmentation results of DeepLab-v3+ and other models on GID dataset. (a) Original image; (b) segmentation result of FCN; (c) segmentation result of U-Net; (d) segmentation result of DeepLab-v3+
Parameter | Value |
---|
base_lr | 0.0001 | lr_decay | 5 | batch_size | 10 | weight_decay | 0.0001 | max_iter | 100 |
|
Table 1. Training parameters
Model | MIoU | Time /h |
---|
FCN | 0.5068 | 1.9 | U-Net | 0.5074 | 3.1 | DeepLab-v3+ | 0.5743 | 1.2 |
|
Table 2. Precision and running time of DeepLab-v3+ and other models
Model | MIoU |
---|
FCN | 0.5606 | U-Net | 0.5782 | DeepLab-v3+ | 0.6426 |
|
Table 3. Precision of DeepLab-v3+ and other models on GID dataset