Infrared and Laser Engineering, Volume. 52, Issue 1, 20220344(2023)
Salient object detection method based on multi-scale feature-fusion guided by edge information
Figures & Tables(13)
Fig. 1. Structural diagram of channel fusion residual block (RCFBlock)
Fig. 3. Structural diagram of expanded spatial attention module guided by edge information (EGSAM)
Table 1. Comparison of the experimental results
View tableView in ArticleTable 1. Comparison of the experimental results
ECSSD PASCAL-S HKU-IS DUTS-TE ${{F} }_{{\beta } }$ $ {M}{A}{E} $ ${{S} }_{{\alpha } }$ ${{F} }_{{\beta } }$ $ {M}{A}{E} $ ${{S} }_{{\alpha } }$ ${{F} }_{{\beta } }$ $ {M}{A}{E} $ ${{S} }_{{\alpha } }$ ${{F} }_{{\beta } }$ $ {M}{A}{E} $ ${{S} }_{{\alpha } }$ MDF 0.832 0.105 0.776 0.768 0.146 0.692 0.861 0.129 0.810 0.730 0.094 0.792 PiCaNet 0.886 0.045 0.917 0.856 0.078 0.848 0.870 0.043 0.904 0.759 0.051 0.869 AFNet 0.908 0.042 0.913 0.821 0.070 0.844 0.888 0.036 0.905 0.792 0.046 0.867 ${ {{\rm{R}}} }^{3}{{\rm{N}}}{{\rm{e}}}{{\rm{t}}}$ 0.914 0.040 0.910 0.845 0.094 0.800 0.893 0.036 0.895 0.785 0.057 0.834 PoolNet 0.915 0.039 0.921 0.822 0.074 0.845 0.892 0.034 0.911 0.809 0.040 0.883 BPFINet 0.928 0.034 0.926 0.845 0.065 0.857 0.911 0.028 0.918 0.838 0.038 0.882 Proposed 0.943 0.033 0.926 0.868 0.069 0.856 0.928 0.033 0.912 0.852 0.037 0.883
Table 2. Parameter quantity and real-time evaluation of EGFMNet
View tableView in ArticleTable 2. Parameter quantity and real-time evaluation of EGFMNet
${\rm{ Parameters}}$ ${\rm{Runtime}/s }$ ${\rm{Frame\;rate}/FPS }$ R3Net 56 156 126 0.030 33 PoolNet 71 383 577 0.033 30 BPFINet 68 326 853 0.033 30 Proposed 60 638 928 0.031 32
Table 3. Ablation experimental results. GA is the basic network trained with BCE Loss, GB is the basic network trained with mixed loss, GC is the complete network with three-stage EGSAM module and trained with BCE Loss, and GD is the complete network with three-stage EGSAM module and trained with mixed loss
View tableView in ArticleTable 3. Ablation experimental results. GA is the basic network trained with BCE Loss, GB is the basic network trained with mixed loss, GC is the complete network with three-stage EGSAM module and trained with BCE Loss, and GD is the complete network with three-stage EGSAM module and trained with mixed loss
Groups Structure Loss Fβ MAE Sα GA Baseline BCE 0.923 0.041 0.908 GB Baseline BCE+BL 0.926 0.040 0.911 GC Baseline+EGSAM(3 stages) BCE 0.936 0.036 0.919 GD Baseline+EGSAM(3 stages) BCE+BL 0.943 0.033 0.926
Table 4. RCFBlock stack quantity verification experiment
View tableView in ArticleTable 4. RCFBlock stack quantity verification experiment
No. Fβ MAE Sα Parameters 1 0.889 0.089 0.831 64 609 584 2 0.943 0.033 0.926 60 638 928 3 0.925 0.040 0.911 48 277 712
Table 5. Verify the experimental results at EGSAM module level
View tableView in ArticleTable 5. Verify the experimental results at EGSAM module level
No. Stage with EGSAM Fβ MAE Sα Parameters Size/MB 1 Baseline 0.923 0.041 0.908 54 437 157 207.67 2 Stage 1 0.936 0.038 0.916 54 734 330 208.79 3 Stage 1+2 0.941 0.035 0.923 55 916 197 213.30 4 Stage 1+2+3 0.943 0.033 0.926 60 638 928 231.32 5 Stage 1+2+3+4 0.942 0.033 0.924 79 521 275 303.35
Table 6. EGSAM fusion coefficient setting experiment
View tableView in ArticleTable 6. EGSAM fusion coefficient setting experiment
No. α Fβ MAE Sα 1 Baseline 0.923 0.041 0.908 2 1 0.842 0.104 0.832 3 0.1 0.911 0.052 0.894 4 0.05 0.926 0.041 0.913 5 0.01 0.943 0.033 0.926 6 0.005 0.940 0.037 0.920
Tools
Get Citation
Copy Citation Text
Xiangjun Wang, Mingyang Li, Lin Wang, Feng Liu, Wei Wang. Salient object detection method based on multi-scale feature-fusion guided by edge information[J]. Infrared and Laser Engineering, 2023, 52(1): 20220344
Paper Information
Category: Image processing
Received: May. 20, 2022
Accepted: --
Published Online: Feb. 9, 2023
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20