Laser & Optoelectronics Progress, Volume. 60, Issue 6, 0610015(2023)
Design of Shoe Print Feature Extraction Network Integrating Global and Local Features
Figures & Tables(14)
Fig. 2. Structure of improving MBConv block. (a) Original MBConv block; (b) MBConv block's improvement
Fig. 6. Presentation of train dataset. (a) Random erase; (b) Gaussian noise; (c) random rotate; (d) random resized crop
Fig. 7. CMC curves of different attention mechanisms and layer feature fusion output on CSS-200
Fig. 8. CMC curves of whether to use fast normalized fusion. (a) CSS-200; (b) CS-Database; (c) FID-300
Table 1. Structure of EfficientNet-B3
View tableTable 1. Structure of EfficientNet-B3
Stage Operator Resolution（H×W） Channels Layers 1 Conv3×3 300×300 40 1 2 MBConv1，k3×3 150×150 24 2 3 MBConv6，k3×3 75×75 32 3 4 MBConv6，k5×5 38×38 48 3 5 MBConv6，k3×3 19×19 96 5 6 MBConv6，k5×5 19×19 136 5 7 MBConv6，k5×5 10×10 232 6 8 MBConv6，k3×3 10×10 384 2 9 Conv1×1 & GMP & FC 10×10 1536 1
Table 2. Composition of different data sets
View tableTable 2. Composition of different data sets
Dataset Query Gallery Negative CS-Database（High-quality） 100 100 CS-Database（Blood） 53 9 91 CS-Database（Blood enhancement） 53 9 91 CS-Database（Dust） 66 9 91 FID-300 300 300 875 CSS-200 200 200 4800
Table 3. Comparison of proposed method and other methods on CS-Database dataset
View tableTable 3. Comparison of proposed method and other methods on CS-Database dataset
Method Accuracy /% top1% top2% top3% top4% top5% High-quality POC 99 99 100 100 100 MCNCC 99 100 100 100 100 Local Semantic Patch and Manifold Ranking 99 100 100 100 100 hybrid features and neighboring images 99 100 100 100 100 Local Semantic Filter Bank 99 100 100 100 100 SCDA 100 100 100 100 100 Proposed method 100 100 100 100 100 Blood POC 45.3 50.9 54.7 64.2 66 MCNCC 92.5 98.1 100 100 100 Local Semantic Patch and Manifold Ranking 79.2 94.3 94.3 94.3 96.2 hybrid features and neighboring images 92.5 100 100 100 100 Local Semantic Filter Bank 94.3 100 100 100 100 SCDA 100 100 100 100 100 Proposed method 100 100 100 100 100 Dust POC 47 47 51.5 51.5 54.5 MCNCC 86.4 89.4 90.9 90.9 95.5 Local Semantic Patch and Manifold Ranking 83.3 84.8 87.9 90.9 90.9 hybrid features and neighboring images 89.4 93.9 95.5 95.5 95.5 Local Semantic Filter Bank 95.45 96.97 98.48 98.48 98.48 SCDA 87.87 90.9 90.9 90.9 90.9 Proposed method 95.45 98.48 98.48 98.48 98.48
Table 4. Comparison of proposed method and other methods on CSS-200 dataset
View tableTable 4. Comparison of proposed method and other methods on CSS-200 dataset
Method Accuracy /% Params FLOPs top1% top2% top5% top10% Conv-25088 62.5 85.5 92.5 94 116M 1.6×1010 SCDA 92.5 94.5 95 96.5 138M 1.9×1010 Resnet101 59.5 80.5 82.5 91.5 49M 7.6×109 ResneXt 64 87 91 93.5 53M 4.2×109 Resnet50-FPN 78 85.5 90 91.5 60M 9.1×109 SENet 79.5 85 85 93.5 146M 2.1×1010 Proposed method 94.5 96.5 98 100 23M 2.6×109
Table 5. Comparison of proposed method and other methods on FID-300 dataset
View tableTable 5. Comparison of proposed method and other methods on FID-300 dataset
Method Accuracy /% top1% top10% Local Semantic Filter Bank 73 93.7 Local Semantic Patch and Manifold Ranking 61.3 88 hybrid features and neighboring images 71.8 87.3 MCNCC 79 89 CABM 58 79 Conv-25088 47.16 76.59 SCDA 50.84 79.93 Proposed method 69.8 90.94
Tools
Get Citation
Copy Citation Text
Yiran Xin, Yunqi Tang, Nengbin Cai. Design of Shoe Print Feature Extraction Network Integrating Global and Local Features[J]. Laser & Optoelectronics Progress, 2023, 60(6): 0610015
Paper Information
Category: Image Processing
Received: Dec. 22, 2021
Accepted: Jan. 27, 2022
Published Online: Mar. 16, 2023
The Author Email: Yunqi Tang (tangyunqi@ppsuc.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20