• Laser & Optoelectronics Progress
  • Vol. 57, Issue 24, 241023 (2020)
Qi Cheng, Guodong Wang*, and Yi Zhao
Author Affiliations
  • College of Computer Science and Technology, Qingdao University, Qingdao, Shandong 266071, China
  • show less
    DOI: 10.3788/LOP57.241023 Cite this Article Set citation alerts
    Qi Cheng, Guodong Wang, Yi Zhao. Text Detection Based on Split-Attention and Path Enhancement Feature Pyramid[J]. Laser & Optoelectronics Progress, 2020, 57(24): 241023 Copy Citation Text show less
    Detection results of the two algorithms. (a) Algorithm based on bounding box; (b) algorithm based on segmentation
    Fig. 1. Detection results of the two algorithms. (a) Algorithm based on bounding box; (b) algorithm based on segmentation
    Overall structure of the network. (a) Backbone; (b) PEFPN; (c) post processing algorithm
    Fig. 2. Overall structure of the network. (a) Backbone; (b) PEFPN; (c) post processing algorithm
    Structure of PEFPN. (a) Top-down; (b) bottom-up; (c) feature fusion
    Fig. 3. Structure of PEFPN. (a) Top-down; (b) bottom-up; (c) feature fusion
    Generation process of label. (a) Calculation of compression mask; (b) real mask; (c) compression mask
    Fig. 4. Generation process of label. (a) Calculation of compression mask; (b) real mask; (c) compression mask
    Effect comparison of verification experiment. (a) VGG+FPN; (b) VGG+PEFPN; (c) Resnet-50+FPN;
    Fig. 5. Effect comparison of verification experiment. (a) VGG+FPN; (b) VGG+PEFPN; (c) Resnet-50+FPN;
    Visualization of network feature maps. (a) Feature map; (b) binary map; (c) result map
    Fig. 6. Visualization of network feature maps. (a) Feature map; (b) binary map; (c) result map
    Detection results on different data sets
    Fig. 7. Detection results on different data sets
    BackboneICDAR2015
    FPNPEFPN
    PRFmeanPRFmean
    VGG[35]74.369.671.976.572.374.3
    ResNet-50[1]80.475.978.181.276.478.7
    ResNet-101[1]82.176.779.382.578.380.3
    SENet[36]81.677.379.482.478.180.2
    ResNeSt-50[25]82.579.481.083.180.381.7
    ResNeSt-101[25]83.080.881.983.881.782.7
    Table 1. Result of verification experiment unit: %
    MethodCTW1500Total-TextICDAR2015MSRA-TD500
    PRFmeanPRFmeanPRFmeanPRFmean
    CTPN[37]60.453.856.9------74.251.660.9------
    SegLink[38]42.340.040.830.323.826.773.176.875.086.070.077.0
    EAST[39]78.749.160.450.036.242.083.673.578.287.367.476.1
    TextSnake[40]67.985.375.682.774.578.484.980.482.682.774.578.4
    PixeLink[20]------------82.981.782.383.073.277.8
    PSENet[22]80.675.678.081.875.178.381.579.780.6------
    Our(50)80.975.978.382.074.778.283.180.381.783.574.879.0
    Our (101)81.376.278.782.575.779.083.881.782.784.075.179.3
    Table 2. Test results of different algorithms on multiple data sets unit: %
    Qi Cheng, Guodong Wang, Yi Zhao. Text Detection Based on Split-Attention and Path Enhancement Feature Pyramid[J]. Laser & Optoelectronics Progress, 2020, 57(24): 241023
    Download Citation