图共 9个 表共 3
    • 图  1  整体算法流程

      Figure 1.  Overall algorithm flow

    • 图  2  镜头突变前后的提取效果对比

      Figure 2.  Comparison of extraction effect before and after lens mutation

    • 图  3  不同$ {t}_{H} $下算法的表现

      Figure 3.  Algorithm performance under different $ {t}_{H} $

    • 图  4  提取效果对比

      Figure 4.  Comparison of extraction effect

    • 图  5  R-SqueezeNet结构

      Figure 5.  Structure of R-SqueezeNet

    • 图  6  Fire module计算流程

      Figure 6.  Calculation process of fire module

    • 图  7  残差结构

      Figure 7.  Residual structure

    • 图  8  不同前景调整方式的分类精度对比

      Figure 8.  Comparison of different foreground resizing method

    • 图  9  不同$ {t}_{H} $下的算法性能对比

      Figure 9.  Performance comparison of algorithms at different $ {t}_{H} $

    • ModelNumber of Fire moduleAccuracy/%Inference time/msSize/MB
      This paper cats vs dogs cifar-10
      SqueezeNet190.5689.991.61.340.16
      294.3590.6193.851.70.33
      396.3292.9695.742.30.89
      496.5893.1796.133.781.5
      596.6793.3996.374.042.8
      R-SqueezeNet396.5593.296.12.880.89
      496.7393.3296.353.921.5
      596.893.4996.554.562.8
      697.0193.5696.617.024.1
      797.293.7196.697.686.4
      897.3893.8796.739.38.9

      表 1  分类网络模型对比

      Table 1.  Comparison of classification network models

    • BasedBackboneFD/%MD/%
      Non-adaptive extractionLeNet13.88.2
      AlexNet9.13.3
      ZFNet8.72.9
      R-SqueezeNet8.62.7
      Adaptive extractionLeNet10.38.2
      AlexNet5.43.3
      ZFNet5.12.9
      R-SqueezeNet4.92.7

      表 2  基于自适应和非自适应前景提取的算法对比

      Table 2.  Comparison of algorithms based on adaptive and non-adaptive foreground extraction

    • AlgorithmBackboneSize/MBFD/%MD/%FPS
      SSD[6]VGG16[25]95.75.32.91
      RetinaNet[7]ResNet50[19]146.14.22.4<1
      YOLOv2[8]Darknet19[8]1944.52.5<1
      YOLOv3[9]Darknet53[9]246.94.22.31
      YOLOv3-tiny[9]Darknet13[9]35.67.63.75
      This paperR-SqueezeNet0.894.92.744

      表 3  本文算法和传统目标检测算法对比

      Table 3.  Comparison with traditional object detection algorithm