• Laser & Optoelectronics Progress
  • Vol. 59, Issue 2, 0215004 (2022)
Jiajun Zhang, Yunqi Tang*, Zhixiong Yang, and Pengzhi Geng
Author Affiliations
  • School of Investigation, People's Public Security University of China, Beijing 100038, China
  • show less
    DOI: 10.3788/LOP202259.0215004 Cite this Article Set citation alerts
    Jiajun Zhang, Yunqi Tang, Zhixiong Yang, Pengzhi Geng. Shoe Type Recognition Algorithm Based on Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0215004 Copy Citation Text show less
    Residual structure of the ResNet50
    Fig. 1. Residual structure of the ResNet50
    Structure of the attention mechanism module
    Fig. 2. Structure of the attention mechanism module
    Residual structure after introducing attention module
    Fig. 3. Residual structure after introducing attention module
    Algorithm flow chart after introducing attention module
    Fig. 4. Algorithm flow chart after introducing attention module
    Equipment construction and walking route
    Fig. 5. Equipment construction and walking route
    Schematic diagram of capturing shoe images
    Fig. 6. Schematic diagram of capturing shoe images
    Schematic diagram of correcting shoe images
    Fig. 7. Schematic diagram of correcting shoe images
    Recognition accuracy curves of different attention mechanism models
    Fig. 8. Recognition accuracy curves of different attention mechanism models
    Visualization results of different models. (a) Original image; (b) Grad-CAM; (c) Guided Grad-CAM
    Fig. 9. Visualization results of different models. (a) Original image; (b) Grad-CAM; (c) Guided Grad-CAM
    Comparison of different feature layers and convolution feature aggregation methods
    Fig. 10. Comparison of different feature layers and convolution feature aggregation methods
    ModelRank-1mAP
    ResNet5061.0946.34
    ResNet50+SENet59.1443.79

    ResNet50+CBAM

    ResNet50+ECA-Net

    Ours

    62.65

    59.14

    65.37

    45.86

    45.71

    47.99

    Table 1. Recognition accuracy of different attention mechanism models

    Layer and feature

    aggregation method

    Rank-1mAP
    FC65.3747.99
    Layer 4+AvgPool69.6552.59

    Layer 4+AvgPool+MaxPool

    Layer 4+MaxPool

    73.15

    73.93

    54.44

    54.87

    Table 2. Comparison of different feature layers and convolution feature aggregation methods
    εRank-1mAP
    Without Label Smoothing73.9354.87
    0.1270.0455.80
    0.1371.2156.65
    0.1474.3256.97
    0.1569.6555.84
    Table 3. Influence of hyper parameter ε on recognition accuracy
    Jiajun Zhang, Yunqi Tang, Zhixiong Yang, Pengzhi Geng. Shoe Type Recognition Algorithm Based on Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0215004
    Download Citation