• Infrared and Laser Engineering
  • Vol. 50, Issue 8, 20200510 (2021)
Weipeng Li, Xiaogang Yang, Chuanxiang Li, Ruitao Lu, Xueli Xie, and Chuan He
Author Affiliations
  • Institute of Missile Engineering, Rocket Force Engineering University, Xi’an 710025, China
  • show less
    DOI: 10.3788/IRLA20200510 Cite this Article
    Weipeng Li, Xiaogang Yang, Chuanxiang Li, Ruitao Lu, Xueli Xie, Chuan He. Infrared object detection network compression using Lp normalized weight[J]. Infrared and Laser Engineering, 2021, 50(8): 20200510 Copy Citation Text show less
    Weight distribution of a neuron with respect to p
    Fig. 1. Weight distribution of a neuron with respect to p
    Sparsity of weight with respect to p at convolutional layers
    Fig. 2. Sparsity of weight with respect to p at convolutional layers
    Training process of sparse neural network for object detection
    Fig. 3. Training process of sparse neural network for object detection
    Result comparison of infrared object detection between classical neural networks and sparse neural networks
    Fig. 4. Result comparison of infrared object detection between classical neural networks and sparse neural networks
    [in Chinese]
    Fig. 5. [in Chinese]
    Comparison of convergence process between SGD and LpSGD
    Fig. 5. Comparison of convergence process between SGD and LpSGD
    ClassificationTrainingTestTotal
    Class 120828236
    Class 221026236
    Class 321930249
    Class 419229221
    Total829113942
    Table 1. Simulated infrared dataset
    MethodScaleAPmAP
    BackboneDetectorClass 1Class 2Class 3Class 4
    Faster R-CNNDense26 852 41614 511 1400.9120.8850.9270.9720.925
    Sparse5 337 35214 511 1300.9100.8750.9360.9820.926
    SSD300Dense22 943 9361 202 9580.8930.8790.9140.9650.914
    Sparse4 103 3961 202 9580.8890.8670.9240.9810.917
    YOLOv3Dense55 294 6886 245 1960.9140.8980.9190.9720.926
    Sparse14 829 7426 245 1960.9060.8950.9270.9840.928
    Table 2. Object detection model and result on simulated infrared dataset
    MethodFaster R-CNNSSD 300YOLOv3
    DenseSparseDenseSparseDenseSparse
    Nonzero parametersBackbone26 852 41615 756 21622 943 93614 995 95255 294 68837 291 638
    Detector14 593 14014 593 1403 341 5503 341 5506 331 3576 331 357
    APAero0.8330.8260.8540.8470.8010.802
    Bike0.7810.7730.7980.7950.8480.845
    Bird0.7350.7370.7020.7120.7160.726
    Boat0.5320.5280.5680.5430.6520.641
    Bottle0.4870.4930.4570.4740.6380.647
    Bus0.7740.7650.7900.7810.8610.858
    Car0.7450.7480.7570.7520.8580.859
    Cat0.8870.8720.7560.7650.8470.857
    Chair0.4490.4430.8710.8650.5470.541
    Cow0.7650.7710.5240.5420.7150.725
    Table0.5480.5360.7680.7640.6900.681
    Dog0.8650.8570.6050.6120.8280.827
    Horse0.8170.8250.8680.8740.8420.846
    Mbike0.8040.7980.8240.8460.8210.831
    Person0.7940.7820.8200.8110.8070.802
    Plant0.3910.3870.4580.4470.4410.437
    Sheep0.7230.7250.7520.7470.6960.688
    Sofa0.6080.5950.6910.6980.6990.696
    Train0.8090.8140.8090.8120.8250.834
    Tv0.6120.6070.6720.6670.7180.722
    mAP0.6980.6940.7170.7180.7420.743
    Table 3. Object detection model and result on VOC2007 dataset
    Weipeng Li, Xiaogang Yang, Chuanxiang Li, Ruitao Lu, Xueli Xie, Chuan He. Infrared object detection network compression using Lp normalized weight[J]. Infrared and Laser Engineering, 2021, 50(8): 20200510
    Download Citation