• Laser & Optoelectronics Progress
  • Vol. 60, Issue 10, 1010014 (2023)
Yingfeng Zhou, Rongfen Zhang, Yuhong Liu*, and Kuan Li
Author Affiliations
  • College of Big Data and Information Engineering, Guizhou University, Guiyang 550025, Guizhou, China
  • show less
    DOI: 10.3788/LOP213356 Cite this Article Set citation alerts
    Yingfeng Zhou, Rongfen Zhang, Yuhong Liu, Kuan Li. Marine Fish Detection Algorithm Based on RetinaNet[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1010014 Copy Citation Text show less
    Structure of improved RetinaNet
    Fig. 1. Structure of improved RetinaNet
    Structure of DenseNet
    Fig. 2. Structure of DenseNet
    Network structure of dense block
    Fig. 3. Network structure of dense block
    Structure of CBAM
    Fig. 4. Structure of CBAM
    Structure of CAM
    Fig. 5. Structure of CAM
    Structure of SAM
    Fig. 6. Structure of SAM
    Structure of PFPN
    Fig. 7. Structure of PFPN
    Sample diagram of dataset
    Fig. 8. Sample diagram of dataset
    Loss value change curve
    Fig. 9. Loss value change curve
    Comparison of accuracy of each category
    Fig. 10. Comparison of accuracy of each category
    Comparison of detection results. (a) Detection result of original algorithm; (b) detection result of improved algorithm
    Fig. 11. Comparison of detection results. (a) Detection result of original algorithm; (b) detection result of improved algorithm
    LayerSettingOutput size
    Convolution7*7 conv,stride 2112*112
    Pooling3*3 MaxPool,stride 256*56
    Dense block 1[1*1 conv,3*3 conv]*656*56
    Transition layer 11*1 conv,2*2 AvgPool,stride 228*28
    Dense block 2[1*1 conv,3*3 conv]*1228*28
    Transition layer 21*1 conv,2*2 AvgPool,stride 214*14
    Dense block 3[1*1 conv,3*3 conv]*2414*14
    Transition layer 31*1 conv,2*2 AvgPool,stride 27*7
    Dense block 4[1*1 conv,3*3 conv]*167*7
    Classification layer

    7*7 global AvgPool,

    1000D fully-connected,softmax

    1*1
    Table 1. Structure of DenseNet-121
    CategoryParameter
    CPUIntel(R)Core(TM)i7-7800X
    GPUNvidia GTX 1080Ti

    Operating system

    Development software

    Development framework

    Ubuntu16.04

    PyCharm

    Pytorch

    Table 2. Experimental environment
    AlgorithmPrecision /%mAP /%
    CenterNet84.6584.89
    YOLOV482.1385.77
    Efficientdet86.7287.51
    YOLOV387.5985.79
    SSD84.6786.11
    Ref.[20-85.99
    Ref.[21-89.00
    Proposed92.7892.12
    Table 3. Precision comparison of different algorithms
    GroupMethodmAP /%Speed /(ms·image-1Memory /MBModel size /MB
    0RetinaNet87.41894216242.0
    1DenseNet-12188.75463247165.1
    2DenseNet-121+CBAM91.21513358168.5
    3DenseNet-121+PFPN90.53503471167.7
    4DenseNet-121+soft-NMS88.95483201165.5
    5DenseNet-121+CBAM+PFPN91.75543411168.7
    6DenseNet-121+CBAM+soft-NMS91.62533315168.9
    7DenseNet-121+PFPN+soft-NMS90.94523398168.4
    8Proposed92.12563408166.6
    Table 4. Ablation experiment