• Laser & Optoelectronics Progress
  • Vol. 58, Issue 2, 0215003 (2021)
Tao Zhang and Le Zhang*
Author Affiliations
  • School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
  • show less
    DOI: 10.3788/LOP202158.0215003 Cite this Article Set citation alerts
    Tao Zhang, Le Zhang. Multiscale Feature Fusion-Based Object Detection Algorithm[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0215003 Copy Citation Text show less
    Network structure based on multi-scale feature fusion
    Fig. 1. Network structure based on multi-scale feature fusion
    Network structure of different feature fusion modules. (a) FPN module; (b) LFF module; (c) HFF module
    Fig. 2. Network structure of different feature fusion modules. (a) FPN module; (b) LFF module; (c) HFF module
    Loss function curves before and after adding feature fusion module. (a) Before adding; (b) after adding
    Fig. 3. Loss function curves before and after adding feature fusion module. (a) Before adding; (b) after adding
    Detection results before and after adding feature fusion module. (a) Before adding; (b) after adding
    Fig. 4. Detection results before and after adding feature fusion module. (a) Before adding; (b) after adding
    AlgorithmBackbone networkmAP /%
    SSD300VGG-1669.7
    RetinaNetResNet-5070.2
    Libra RetinaNetResNet-5070.4
    Proposed algorithmResNet-5072.6
    Table 1. mAP values of different detection algorithms on PASCAL VOC dataset
    CategorySSD300RetinaNetLibra RetinaNetProposed algorithm
    Aero75.775.475.477.4
    Bike78.480.279.680.1
    Bird67.272.172.173.5
    Boat64.460.764.867.3
    Bottle38.939.437.943.1
    Bus79.978.979.481.1
    Car83.379.079.080.3
    Cat84.686.485.187.4
    Chair49.552.552.654.9
    Cow67.164.168.074.1
    CategorySSD300RetinaNetLibra RetinaNetProposed algorithm
    Table63.767.766.568.5
    Dog79.280.180.582.2
    Horse80.579.478.981.4
    Mbike79.778.278.179.7
    Person75.274.274.575.6
    Plant37.243.042.345.3
    Sheep69.968.470.470.9
    Sofa68.571.371.372.0
    Train81.683.183.784.1
    TV69.370.368.872.7
    Table 2. AP values of different categories unit: %
    HFF moduleLFF modulemAP /%
    70.4
    72.0
    72.2
    72.6
    Table 3. Results of ablation study
    CategoryLFF moduleHFF module
    Aero76.577.9
    Bike80.280.5
    Bird74.172.6
    Boat66.066.6
    Bottle40.741.9
    Bus78.782.4
    Car80.279.4
    Cat86.186.8
    Chair54.953.2
    Cow70.969.2
    Table70.268.3
    Dog82.282.2
    Horse82.582.5
    CategoryLFF moduleHFF module
    Mbike79.280.5
    Person75.075.0
    Plant44.745.0
    Sheep69.870.7
    Sofa73.573.3
    Train83.983.4
    TV70.571.6
    Table 4. AP values of each category after ablation learning unit: %
    AlgorithmBackboneAPAP50AP75APSAPMAPL
    SSD512VGG-1625.744.126.69.229.039.0
    RetinaNetResNet-5035.655.638.120.839.546.1
    Libra RetinaNetResNet-5037.556.939.922.441.449.2
    Proposed algorithmResNet-5038.858.541.322.942.650.4
    RetinaNetResNet-10137.857.540.820.942.149.6
    Libra RetinaNetResNet-10139.158.641.722.643.851.4
    Proposed algorithmResNet-10140.359.942.923.144.853.3
    Table 5. AP values of different detection algorithms on MSCOCO dataset unit: %
    HFF moduleLFF moduleAPAP50AP75APSAPMAPL
    37.556.939.922.441.449.2
    38.558.341.022.442.650.2
    38.558.440.923.842.449.7
    38.858.541.322.942.650.4
    Table 6. Results after ablation learning on COCO val-2017 unit: %
    Tao Zhang, Le Zhang. Multiscale Feature Fusion-Based Object Detection Algorithm[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0215003
    Download Citation