Fig 1.
The W-Net network structure consists of two modules: Two-path convolution module and focus on prominent areas module.
Fig 2.
U-Net has poor ability to handle remote sensing images with excessive contextual information.
DUNet obtains better segmentation results by deepening the network structure.
Fig 3.
The two-path convolution module contains two independent paths of FCN and DUNet for feature extraction of the image, and the fused image retains more feature information.
Fig 4.
Accuracy (PA) vs. precision (mIoU) on the training set.
Table 1.
The accuracy of each model in object segmentation and the mIoU results are shown on the training set by qualitative experiments.
The best results in the experiments are indicated by the values in bold in each column.
Table 2.
Accuracy (PA), precision (mIoU), and consistency (Kappa) obtained by testing using different backbones in the fully convolutional neural network (FCN).
Table 3.
Data comparison of the proposed DUNet network modules in terms of mIoU, PA, and Kappa.
Table 4.
In contrast to the use of FCN, U-Net, and DUNet methods for determining dual paths, verifying that image classification (cls) must be applied for the image segmentation task.
The values in bold font indicate the best type of dual path convolution module.
Table 5.
Experimental results of the optimal solution for the reduction ratio.
Fig 5.
mIoU (%) for W-Net optimization using differentbase_lr.
Table 6.
Semantic segmentation of the HSR remote sensing dataset iSAID dataset using state-of-the-art methods to compare the experimental mIoU.
Table 7.
The bold values in each column indicate the IoU best results for each type of object segmentation on the iSAID dataset.
Each abbreviation is explained as: SV(Small vehicle), BD(Baseball diamond), HC(Helicopter), SP(Swimming pool), TC(Tennis court), LV(Large vehicle), SC(Storage tank), GTF(Ground field track), SBF(Soccer-ball field), BC(Basketball court), and RA(Roundabout).
Fig 6.
Qualitative experiments based on the use of the W-Net network on the iSAID test set were compared with three other methods.
(a) iSAID test set image, (b) W-Net network object segmentation map, (c) U2-Net network object segmentation map, (d) FCN network object segmentation map, (e) U-Net network object segmentation map. To facilitate visualization of the mapping relationships of the objects of the experimental results, we resize the images. Legend: Scene 1 (tennis court), Scene 2 (airplane), Scene 3 (small vehicle, large vehicle), Scene 4 (small vehicle, port, ship).