Otwarty dostęp

Research on Deep Learning-based Image Processing and Classification Techniques for Complex Networks

Jiangli Liu

Jiangli Liu

Business School, Guilin University of Electronic TechnologyGuilin, China
Wyszukaj tego autora
Sciendo|Google Scholar
Liu, Jiangli
, 
Jinfeng Li

Jinfeng Li

Business School, Guilin University of Electronic TechnologyGuilin, China
Wyszukaj tego autora
Sciendo|Google Scholar
Li, Jinfeng
 oraz 
Guangyan Huang

Guangyan Huang

Business School, Guilin University of Electronic TechnologyGuilin, China
Wyszukaj tego autora
Sciendo|Google Scholar
Huang, Guangyan
  
17 mar 2025
Research on Deep Learning-based Image Processing and Classification Techniques for Complex Networks's Cover Image
O artykule
Poprzedni artykuł
Następny artykuł
Zacytuj
Pobierz okładkę
Data publikacji: 17 mar 2025
Otrzymano: 21 paź 2024
Przyjęty: 05 lut 2025
DOI: https://doi.org/10.2478/amns-2025-0351
Słowa kluczowe
© 2025 Jiangli Liu et al., published by Sciendo
This work is licensed under the Creative Commons Attribution 4.0 International License.

Figure 1.

The flow chart of image texture feature extraction algorithm
The flow chart of image texture feature extraction algorithm

Figure 2.

Structure of decanet network model
Structure of decanet network model

Figure 3.

Dsa_aspp module
Dsa_aspp module

Figure 4.

Linear interpolation
Linear interpolation

Figure 5.

Network parameter statistics profile
Network parameter statistics profile

Figure 6.

Network parameter statistics profile
Network parameter statistics profile

Figure 7.

Training loss and test accuracy curve
Training loss and test accuracy curve

Comparison Experimen results on the Cityscapes dataset

Model γ Backbone MIoU-val(%) MIoU-test(%) Fps
ENet 0.4 - - 56 74.5
ICNet 0.9 PSPNet50 66.4 63.8 25.9
BiSeNetV1 0.65 Xception-39 67.9 67.1 106.9
BiSeNetV1-L 0.65 ResNet-18 71.6 69 64.7
BiSeNetV2 0.4 - 71.8 70.4 144.1
BiSeNetV2-L 0.4 - 72.7 70.3 43.4
BiSeNetV3-1 0.4 STDC1 69.6 67.9 243.8
BiSeNetV3-2 0.4 STDC2 71.4 70.3 167.3
PreactResNet1 0.4 STDC1 71.1 69.6 255.8
PreactResNet2 0.4 STDC2 74.7 73.6 184.3

Comparison results of model performance

Model Underlying network 1-Stage Stanford Dogs (%) CUB-200-2011 (%)
ResNet50 ResNet50 √ 88.7 85.7
GP-256 VGG16 × 89.1 87
MaxEnt DenseNet161 √ 89.6 87.8
DFL-CNN ResNet50 √ 93.7 88.6
NTS-Net ResNet50 √ 94.2 88.7
Cross-X ResNet50 × 94.9 88.9
CIN ResNet101 √ 93.6 89.3
ACNet ResNet50 √ 93.4 89.3
S3N ResNet50 √ 93.1 89.7
FDL ResNet161 √ 90.9 90.3
PMG ResNet50 √ 3.5 90.8
FBSD ResNet161 √ 94.1 91
API-Net ResNet161 √ 96.3 91.2
StackedLSTM GoogleNet √ 3.5 91.6
CAL ResNet101 √ 94.7 91.8
HDML GoogleNet √ 95.3 92.4
DCML ResNet50 √ 95.9 92.8
ViT ViT-B_16 √ 15.8 91.6
TransFG ViT-B_16 √ 96.4 92.6
FFVT ViT-B_16 √ 96.4 92.6
RAMS-Trans ViT-B_16 √ 96.7 92.7
AFTrans ViT-B_16 √ 6.6 92.8
PreactResNet ViT-B_16 √ 97 93

Comparison Experimen results on the CamVid dataset

Model Backbone Resolution MIoU(%) Fps
ENet - 940×710 48 55
ICNet PSPNet50 940×710 65.8 27.1
BiSeNetV1 Xception-39 940×710 62.1 177.1
BiSeNetV1-L ResNet-18 940×710 66.8 113.9
BiSeNetV2 - 940×710 70.5 122.7
BiSeNetV2-L - 940×710 70.8 41.9
BiSeNetV3-1 STDC1 940×710 70.6 196
BiSeNetV3-2 STDC2 940×710 71.2 153.4
PreactResNet STDC1 940×710 69.6 221.5
PreactResNet STDC2 940×710 75.4 143.2
Język:
Angielski
Częstotliwość wydawania:
1 razy w roku
Dziedziny czasopisma:
Nauki biologiczne, Nauki biologiczne, inne, Matematyka, Matematyka stosowana, Matematyka ogólna, Fizyka, Fizyka, inne
Kanał RSS czasopisma