64
S. Hu et al.
Table 1. Result of experience.
Segmentation model IoU Dice
FCN-8s
0.653 0.769
FCN-8s and attention 0.675 0.796
U-net
0.762 0.857
U-net and attention
0.784 0.879
Fig. 3. Experimental result.
attention mechanism has a better segmentation effect, and boundary is clearer.
The labeling part of the resulting figure can better reflect that the model combined with attention mechanism. By comparing the segmentation effects of every
image, our model effectively reduces the mistakes of dividing roads into buildings, and the degree of blurring of the boundary; it is more accurate for building
shape segmentation. However, the method can be further improved due to some
segmentation mistakes. For example, when dataset ratio is 21:5 and the segmentation network is U-net(third column), our model has a clearer segmentation
boundary than baseline network, but by comparing it with ground truth, our
model marks the road with similar building pixel values as building, resulting
in segmentation mistakes. In general, the results prove that the network model
combined with multi-scale and attention mechanism has a better segmentation
effect and clearer architectural detail segmentation.