• Laser & Optoelectronics Progress
  • Vol. 58, Issue 24, 2410002 (2021)
Ming Yu, Jijun Zhang, Yingchun Guo*, Meng Zhang, and Dan Wang
Author Affiliations
  • School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China
  • show less
    DOI: 10.3788/LOP202158.2410002 Cite this Article Set citation alerts
    Ming Yu, Jijun Zhang, Yingchun Guo, Meng Zhang, Dan Wang. Image Aesthetics Retargeting Algorith Based on Multi-Level Attention Fusion[J]. Laser & Optoelectronics Progress, 2021, 58(24): 2410002 Copy Citation Text show less
    Flowchart of image aesthetic retargeting based on attention fusion
    Fig. 1. Flowchart of image aesthetic retargeting based on attention fusion
    Fusion of importance maps. (a) Original images; (b) aesthetic feature maps; (c) Esal; (d) fusions of Egrad ; (e) importance maps
    Fig. 2. Fusion of importance maps. (a) Original images; (b) aesthetic feature maps; (c) Esal; (d) fusions of Egrad ; (e) importance maps
    Image retargeting process based on the energy transfer
    Fig. 3. Image retargeting process based on the energy transfer
    Seven retargeting results for different type of images. (a) With a single large subject; (b) with a complex background; (c) with multiple subjects; (d) with complex subject objects
    Fig. 4. Seven retargeting results for different type of images. (a) With a single large subject; (b) with a complex background; (c) with multiple subjects; (d) with complex subject objects
    Results of ablation experiments. (a) Original images; (b) important map + gradient and straight-line map; (c) aesthetic feature map + important map + gradient and straight-line map
    Fig. 5. Results of ablation experiments. (a) Original images; (b) important map + gradient and straight-line map; (c) aesthetic feature map + important map + gradient and straight-line map
    MethodSRCCPLCCAccuracy /%
    Ref. [19]--66.70
    Ref. [20]--71.42
    Ref. [31]--74.46
    Ref. [32]--75.76
    Ref. [33]--77.40
    Ref. [34]0.558-77.33
    Ref. [35]--81.70
    Ref. [30]0.6120.63681.51
    Ref. [22]0.7520.75581.61
    Ref. [22]0.7560.75781.72
    Proposed algorithm0.7550.75782.28
    Table 1. Comparison of between proposed method and mainstream methods
    Retargeting algorithmFig. 4(a)Fig. 4(b)Fig. 4(c)Fig. 4(d)
    US0.590.600.680.60
    SC0.630.650.620.61
    SNS0.600.610.680.63
    BSC0.660.650.700.68
    InGAN0.400.610.490.62
    Cycle-IR0.600.680.660.70
    Proposed algorithm0.680.730.700.72
    Table 2. Objective evaluation results of 4 images
    Retargeting algorithmSelect numberSelect rate /%
    US1162.90
    SC2125.30
    SNS40710.18
    BSC48912.22
    InGAN1323.30
    Cycle-IR55913.98
    W/o aesthetic feature map77319.33
    Proposed algorithm131232.80
    Table 3. Subjective evaluation statistics
    Retargeting algorithmAverage score
    US0.60
    SC0.63
    SNS0.63
    BSC0.68
    InGAN0.59
    Cycle-IR0.68
    W/o aesthetic feature map0.68
    Proposed algorithm0.71
    Table 4. Objective evaluation statistics
    Ming Yu, Jijun Zhang, Yingchun Guo, Meng Zhang, Dan Wang. Image Aesthetics Retargeting Algorith Based on Multi-Level Attention Fusion[J]. Laser & Optoelectronics Progress, 2021, 58(24): 2410002
    Download Citation