Otwarty dostęp

Research on Capsule Network Based on Attention Mechanism

Yan Jiao

Yan Jiao

School of Computer Science and Engineering Xi’an Technological UniversityXi’an, China
Wyszukaj tego autora
Sciendo|Google Scholar
Jiao, Yan
, 
Li Zhao

Li Zhao

School of Computer Science and Engineering Xi’an Technological UniversityXi’an, China
Wyszukaj tego autora
Sciendo|Google Scholar
Zhao, Li
 oraz 
Hexin Xu

Hexin Xu

School of Computer Science and Engineering Xi’an Technological UniversityXi’an, China
Wyszukaj tego autora
Sciendo|Google Scholar
Xu, Hexin
  
21 lut 2022
Research on Capsule Network Based on Attention Mechanism's Cover Image
O artykule
Poprzedni artykuł
Następny artykuł
Zacytuj
Pobierz okładkę
Data publikacji: 21 lut 2022
Zakres stron: 1 - 8
DOI: https://doi.org/10.21307/ijanmc-2021-011
Słowa kluczowe
© 2021 Yan Jiao et al; published by Sciendo
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Figure 1.

The original capsule network
The original capsule network

Figure 2.

CBAM Module
CBAM Module

Figure 3.

Channel Attention Module
Channel Attention Module

Figure 4.

Spatial Atttion Module
Spatial Atttion Module

Figure 5.

Capsule network based on attention mechanism
Capsule network based on attention mechanism

j_ijanmc-2021-011_utab_001

Procedure Routing algorithm
1: procedure ROUTING ( u ^ i ∣ j , i , l )
2:     for all capsule i in layer l and capsule j in layer (l + 1):bij ←0
3:     for r iterations do
4:         for all capsule i in layer l : ci ← softmax (bi )
5:         for all capsule j in layer (l +1) : sj ← ∑ i C i j u ^ i | j
6:         for all capsule j in layer (l +1) : vj ← squash (sj )
7:         for all capsule i in layer l and capsule j in layer (l + l)
8: return vj

NETWORK MODEL AND PARAMETERS

Layer Parameters
Conv inputChannel:1; outPutChannel:256 kernel size= 9; stride=1
CBAM Channel:256
PrimaryCaps InputChannel:256;OutputCaps:32*6*6output_dim:8;kernel_size:9,stride:2
MgitCaps Inputeaps:32*6*6;out_put_caps:10

NETWORK MODEL ACCURACY

Net Work Name Routing Number CBAM Module Max Accuracy First time
CapsNetl 1 False 99.70999908% 133
CapsNetl_CBAM 1 True 99.66999817% 100
CapsNet2 2 False 99.65000153% 111
CapsNet2_CBAM 2 True 99.61000061% 42
CapsNet3 3 False 99.69000244% 131
CapsNet3_CBAM 3 True 99.62002324% 82
CapsNet4 4 False 99.59999847% 141
CapsNet4_CBAM 4 True 99.55999756% 86
CapsNet5 5 False 99.65000153% 125
Język:
Angielski
Częstotliwość wydawania:
4 razy w roku
Dziedziny czasopisma:
Informatyka, Informatyka, inne
Kanał RSS czasopisma