Uneingeschränkter Zugang

Multi-task learning based feature extraction method in signal processing of high resolution remote sensing video images

Xinming Fan

Xinming Fan

School of Information Engineering, Yancheng Institute of TechnologyYancheng, China
Suche nach diesem Autor auf
Sciendo|Google Scholar
Fan, Xinming
  
21. März 2025
Multi-task learning based feature extraction method in signal processing of high resolution remote sensing video images's Cover Image
Über diesen Artikel
Vorheriger Artikel
Nächster Artikel
Zitieren
COVER HERUNTERLADEN
Online veröffentlicht: 21. März 2025
Eingereicht: 28. Okt. 2024
Akzeptiert: 06. Feb. 2025
DOI: https://doi.org/10.2478/amns-2025-0695
Schlüsselwörter
© 2025 Xinming Fan, published by Sciendo
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Figure 1.

The parameter sharing method of multi-task learning
The parameter sharing method of multi-task learning

Figure 2.

The structure of the label attention module
The structure of the label attention module

Figure 3.

The loss value of different model training
The loss value of different model training

Figure 4.

Test results on the Vaihingen data set
Test results on the Vaihingen data set

Figure 5.

Test results on the WHDLD data set
Test results on the WHDLD data set

Figure 6.

Test results on the DLRSD data set
Test results on the DLRSD data set

Test results of different models on the WHDLD data set

Models Building Roads Walkway Vegetation Naked soil Water Average (%)
MD TANet 62.61 68.74 49.54 82.01 48.96 93.44 67.55
U-Net 60.41 59.42 42.01 78.71 36.97 90.79 61.39
EANet 53.11 63.63 47.76 79.21 35.15 90.76 61.6
PSPNet 60.12 62.75 46.89 73.29 32.99 92.86 61.54
BiseNetv2 58.35 64.14 48.18 71.91 32.57 91.09 61.04
CCNet 61.51 64.25 45.86 77.95 39.03 93.02 63.6
RefineNet 61.14 63.02 42.01 79.39 37.72 93.29 62.76
EMANet 59.69 62.73 47.87 73.11 37.36 92.18 62.16
Deeplabv3+ 61.57 62.93 46.98 79.62 35.05 92.19 63.06

Test results of different models on the DLRSD data set

Models MD TANet EANet CCNet RefineNet Deeplabv3+
Airplane 68.30 64.65 62.58 62.89 64.76
Naked soil 59.05 53.16 54.51 53.89 53.49
Architecture 71.44 74.51 73.38 72.40 72.39
Car 67.53 62.20 65.23 64.38 65.20
Church 59.33 53.020 51.380 53.110 53.55
Courtyard 84.99 81.28 80.53 81.72 80.22
Quay 54.60 42.82 46.30 46.88 45.73
Field 91.83 93.6 90.93 94.66 93.25
Grass 64.28 66.18 65.11 64.78 64.88
Mobile house 61.07 63.96 62.1 61.06 60.84
Walkway 77.61 76.94 73.11 75.71 76.83
Sand beach 65.71 71.78 62.75 68.89 68.18
Ocean 96.73 93.25 90.79 92.56 91.86
Ship 77.07 65.04 70.23 72.28 70.15
Tanks 71.69 76.44 75.07 73.61 68.31
Tree 74.76 72.60 73.35 69.93 71.01
Other waters 83.44 82.18 81.75 80.98 80.04
Average (%) 72.20 70.21 69.36 69.98 69.45

Test results of different models on the Vaihingen data set

Models Roads Building Dwarf vegetation Tree Car Other Average (%)
MD TANet 85.92 87.81 75.71 78.34 68.89 59.22 75.98
U-Net 80.25 80.34 67.55 72.26 65.75 57.55 70.62
EANet 81.49 83.84 71.53 74.96 66.86 55.16 72.31
PSPNet 80.56 80.15 67.71 69.71 67.62 51.85 69.60
BiseNetv2 82.91 82.51 73.74 73.83 68.51 55.41 72.82
CCNet 82.42 83.57 73.15 73.84 70.47 57.82 73.55
RefineNet 83.14 82.21 71.87 70.28 69.02 58.37 72.48
EMANet 81.51 81.84 69.88 69.37 66.18 55.33 70.69
Deeplabv3+ 83.93 83.54 73.57 74.21 69.52 56.48 73.54

Experimental platform configuration

Name Configuring
Hardware CPU Intel(R) Xeon(R) Gold 5115 CPU@2.40GHz
Memory 128GB
GPU Tesla P100
Display storage 16GB
Software Programming language Python 3.7.15
Compiling environment Pycharm
Anaconda version Conda 4.10.3
Depth learning framework Pytorch 1.9.0
Operating system CentOS 7.7

The evaluation results of different models in the validation set

Model Input size mAP@50 mIoU@50 FPS Parameters FLOPS
U-Net 256*256 73.86 70.62 82.42 10.57 25.23
EANet 256*256 75.88 72.31 94.22 7.93 28.13
PSPNet 256*256 72.14 69.60 73.35 11.49 22.12
BiseNetv2 256*256 74.81 72.82 68.32 13.31 33.49
CCNet 256*256 76.94 73.55 76.22 10.77 53.35
RefineNet 256*256 76.42 72.48 88.65 9.42 40.44
EMANet 256*256 75.78 70.69 73.76 12.07 29.65
Deeplabv3+ 256*256 76.49 73.54 76.21 11.25 35.37
MD TANet 256*256 78.10 75.98 90.72 10.38 26.54
Sprache:
Englisch
Zeitrahmen der Veröffentlichung:
1 Hefte pro Jahr
Fachgebiete der Zeitschrift:
Biologie, Biologie, andere, Mathematik, Angewandte Mathematik, Mathematik, Allgemeines, Physik, Physik, andere
Zeitschrift RSS Feed