Laser & Optoelectronics Progress, Volume. 60, Issue 6, 0610008(2023)
Improved ResNet Image Classification Model Based on Tensor Synthesis Attention
Figures & Tables(19)
Fig. 8. Comparison of classification accuracy of different methods on Gaussian noise images
Fig. 9. Comparison of the classification accuracy of different methods for rotating images
Fig. 10. Comparison of the classification accuracy of different methods for the center cropped image
Table 1. Details of datasets
View tableTable 1. Details of datasets
Dataset Number of data Category Training set Testing set CIFAR-10
CIFAR-100
SVHN
60000
60000
99289
10
100
10
50000
50000
73257
10000
10000
26032
Table 2. Abbreviations of various methods
View tableTable 2. Abbreviations of various methods
Method abbreviation Concrete model None CNN CA CNN+Attention CSD CNN+Synthesizer Dense CFSD CNN+Factorized Synthesizer Dense CFSR CNN+Factorized Synthesizer Random CMS CNN+Mixture Synthesizers CTSA（C） CNN+Tensor Synthetic Attention Channel CTSA（H） CNN+Tensor Synthetic Attention Hight CTSA（W）
CTSAR
CNN+Tensor Synthetic Attention Width
CNN+Tensor Synthetic Attention Random
Table 3. Classification accuracy results of different methods for Gaussian noise images
View tableTable 3. Classification accuracy results of different methods for Gaussian noise images
Method Gaussian noise 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10 None 45.96 34.35 27.80 22.83 20.01 17.86 16.23 15.07 14.45 13.76 CA 47.02 35.83 29.21 24.70 21.49 19.33 17.41 16.51 15.42 14.80 CSD 48.71 37.79 31.54 27.30 23.70 22.23 20.18 18.92 17.45 17.03 CFSR 46.51 37.28 31.69 27.56 24.14 22.60 20.62 19.75 18.25 17.76 CFSD 48.90 37.50 30.42 25.89 22.85 20.70 19.33 17.91 16.87 16.07 CMS 48.20 36.80 29.55 25.13 21.45 19.34 17.74 16.61 15.76 15.26 CTSA（C） 49.29 36.81 31.98 28.11 25.49 23.75 22.39 21.44 20.09 19.54 CTSA（H） 44.66 30.92 23.26 19.58 17.08 15.12 14.04 13.35 12.61 12.26 CTSA（W）
CTSAR
42.20
45.93
30.21
34.02
23.65
27.06
19.92
23.01
17.53
19.91
16.41
17.36
15.18
16.02
14.65
14.93
14.33
13.93
13.64
13.40
Table 4. Classification accuracy results of different methods for rotating pictures
View tableTable 4. Classification accuracy results of different methods for rotating pictures
Method Rotation angle/（°） 30 60 90 120 150 180 210 240 270 300 330 None 37.58 26.51 23.07 21.12 19.40 19.39 18.68 18.89 18.27 17.74 17.40 CA 36.54 36.57 23.27 21.25 19.75 19.44 19.70 18.61 18.19 18.46 17.75 CSD 37.90 38.17 23.24 21.79 19.26 19.45 19.36 18.51 17.55 17.60 17.98 CFSR 37.04 26.86 23.10 20.69 19.72 18.69 18.70 18.27 17.69 17.24 17.08 CFSD 39.27 28.85 23.55 22.85 20.37 19.85 19.95 18.71 18.86 18.34 17.88 CMS 37.91 26.93 22.96 21.74 19.61 18.89 18.76 18.01 18.06 17.62 17.65 CTSA（C） 38.71 38.36 24.51 21.80 20.97 20.36 20.89 19.93 19.86 19.12 20.00 CTSA（H） 42.32 30.44 26.28 24.53 22.17 22.18 22.37 21.83 20.85 20.07 20.16 CTSA（W）
CTSAR
45.53
33.72
38.64
24.70
27.21
22.06
25.20
19.88
22.39
19.00
22.23
18.17
22.03
18.03
21.27
17.22
20.61
17.03
20.01
17.23
19.89
16.79
Table 5. Classification accuracy results of different methods for center cropped images
View tableTable 5. Classification accuracy results of different methods for center cropped images
Method Center crop size/pixel 10 15 20 25 30 None 14.89 20.22 32.68 51.15 61.20 CA 13.20 17.98 28.00 47.41 60.58 CSD 13.79 19.29 30.15 48.85 61.38 CFSR 14.03 19.19 30.23 48.89 61.23 CFSD 13.53 19.03 31.28 51.87 63.26 CMS 15.12 20.12 32.58 51.31 62.16 CTSA（C） 13.74 18.93 32.60 53.56 63.62 CTSA（H） 14.70 20.15 31.51 50.56 61.84 CTSA（W）
CTSAR
15.22
14.03
20.53
19.19
33.23
30.23
52.12
48.89
62.21
61.23
Table 6. Classification results of different methods
View tableTable 6. Classification results of different methods
Method Accuracy /% Recall /% Precision /% F1 /% None 83.52 62.31 52.30 71.59 CA 82.30 71.28 68.46 82.37 CSD 84.23 80.56 85.77 78.38 CFSR 72.94 73.59 83.21 80.97 CFSD 87.71 81.32 77.32 67.98 CMS 65.23 88.57 60.58 56.35 CTSA（C） 91.65 90.20 89.91 86.72 CTSA（H） 90.50 89.30 86.35 84.23 CTSA（W） 89.96 87.23 87.25 83.37 CTSAR 88.67 88.59 89.34 82.56
Table 7. Comparison of classification results of residual network models
View tableTable 7. Comparison of classification results of residual network models
Model Accuracy /% ResNet-18 89.33 ResNet-34 88.76 ResNet-50 86.72 ResNet-101 93.12
Table 8. Comparison of classification accuracy of different RTSA Net models
View tableTable 8. Comparison of classification accuracy of different RTSA Net models
Model Accuracy /% RTSA Net-18 90.30 RTSA Net-34 91.23 RTSA Net-50 92.98 RTSA Net-101 96.12
Table 9. Comparison with other advanced methods for classification accuracy and average test running time
View tableTable 9. Comparison with other advanced methods for classification accuracy and average test running time
Model CIFAR-10 CIFAR-100 SVHN Accuracy /% running time /s Accuracy /% running time /s Accuracy /% running time /s ANODE［ 25 ］60.60 0.0378 — — 83.50 0.0308 Sign-symmetry［ 26 ］80.98 0.0341 52.25 0.0301 89.84 0.0317 Improved GAN［ 27 ］88.17 0.0291 — — 91.89 0.0269 CLS-GAN［ 28 ］91.70 0.0431 — — 94.02 0.0277 NNCLR［ 29 ］93.70 0.0289 79.00 0.0289 — — CCT-6/3x1［ 30 ］95.29 0.0321 77.31 0.0271 — — ResNet56 with reSGHMC［ 31 ］96.10 0.0278 80.14 0.0278 — — RTSA Net-101 96.12 0.0258 81.60 0.0260 96.67 0.0262
Tools
Get Citation
Copy Citation Text
Yunfei Qiu, Jiaxin Zhang, Hai Lan, Jiaxu Zong. Improved ResNet Image Classification Model Based on Tensor Synthesis Attention[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0610008
Paper Information
Category: Image Processing
Received: Oct. 29, 2021
Accepted: Jan. 17, 2022
Published Online: Mar. 7, 2023
The Author Email: Jiaxin Zhang (1595775491@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20