• Laser & Optoelectronics Progress
  • Vol. 59, Issue 10, 1015009 (2022)
Yang Cao*, Li Zhang, Junxi Meng, Qian Song, and Letian Zhang
Author Affiliations
  • College of Electronics and Information, Xi'an Polytechnic University, Xi'an 710600, Shaanxi , China
  • show less
    DOI: 10.3788/LOP202259.1015009 Cite this Article Set citation alerts
    Yang Cao, Li Zhang, Junxi Meng, Qian Song, Letian Zhang. Multi-Target Prohibited Item Recognition Algorithm for X-Ray Security Scene[J]. Laser & Optoelectronics Progress, 2022, 59(10): 1015009 Copy Citation Text show less
    Network structure of the algorithm
    Fig. 1. Network structure of the algorithm
    Residual block based on attention mechanism
    Fig. 2. Residual block based on attention mechanism
    Structure of the FE module
    Fig. 3. Structure of the FE module
    Structure of the ARFE module
    Fig. 4. Structure of the ARFE module
    Standard convolution and transformable convolution. (a) Standard convolution kernel; (b) transformable convolution kernel; (c) special forms of transformable convolution kernel
    Fig. 5. Standard convolution and transformable convolution. (a) Standard convolution kernel; (b) transformable convolution kernel; (c) special forms of transformable convolution kernel
    Learning process of the transformable dilated convolution
    Fig. 6. Learning process of the transformable dilated convolution
    Visualization results of different algorithms on MCDD. (a) Original image; (b) label; (c) U-Net; (d) DeepLabv3; (e) our algorithm
    Fig. 7. Visualization results of different algorithms on MCDD. (a) Original image; (b) label; (c) U-Net; (d) DeepLabv3; (e) our algorithm
    GroupARFEDilation rateMIoU /%
    1×-80.97
    2681.65
    31281.93
    41882.26
    52481.11
    Table 1. Performance of the model with different dilation rate
    ModelBaseline networkMIoU /%FPSModel size /Mb
    Ours 1MobileNetv275.9328.348.97
    Ours 2ResNet3476.8118.1929.31
    Ours 3ResNet10183.149.5750.33
    Ours 4ResNet5082.2616.2134.46
    Table 2. Segmentation performance of models under different feature extraction baseline networks
    AlgorithmTime /msMIoU /%FPS
    U-Net62.6272.1415.97
    FCN-8 s64.9870.8515.39
    PSPNet118.9173.036.41
    DeepLabv3141.8479.285.05
    Ours61.6982.2616.21
    Table 3. Segmentation performance of different algorithms on MCDD
    GroupAlgorithmMIoU /%FPS
    -77.8417.19
    CBAM78.1617.04
    CBAM+ARFE79.2316.45
    CBAM+ARFE+FE80.3716.62
    CBAM+ARFE+FE+ASPP80.8916.13
    CBAM+ARFE+FE+TASPP82.2616.21
    Table 4. Effects of different modules on algorithm performance
    GroupAlgorithmBackbone networkMIoU /%
    FCN-8sVGG-1661.57
    PSPNetResNet10177.62
    DeepLabv3ResNet10178.19
    oursResNet5081.03
    Table 5. Performance of different algorithms on PASCAL VOC 2012 dataset
    Yang Cao, Li Zhang, Junxi Meng, Qian Song, Letian Zhang. Multi-Target Prohibited Item Recognition Algorithm for X-Ray Security Scene[J]. Laser & Optoelectronics Progress, 2022, 59(10): 1015009
    Download Citation