Author Affiliations
1Key Laboratory of Wireless Sensor Network and Communication, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 201800, China2University of Chinese Academy of Sciences, Beijing 100864, Chinashow less
Fig. 1. Structure of improved DeepLabv3+ model
Fig. 2. Residual structure of ResNet
Fig. 3. Residual structure with attention mechanism
Fig. 4. Structure of edge refinement module
Fig. 5. Effect comparison of edge refinement module before and after modification. (a) Original image; (b) before correction; (c) after correction
Fig. 6. Image of water level scale taken. (a) Scale of water level is not clear; (b) calibration cannot be observed
Fig. 7. Effect comparison of different semantic segmentation algorithms on water scale segmentation. (a) Original image; (b) ground truth; (c) U-Net; (d) PSPNet; (e) DeepLabv3+; (f) proposed algorithm
Fig. 8. Horizontal plane detection effect comparison of different algorithms. (a) Original image; (b) algorithm 1; (c) algorithm 2; (d) proposed algorithm
Fig. 9. Test results of algorithm under different extreme environmental conditions. (a) Low light environment; (b) strong light environment; (c) foreign body occlusion; (d) blurred water level scales images; (e) calm and clear reflection of water; (f) nighttime water level scale images
Algorithm | mIoU /% | IoU /% | MPA /% |
---|
SegNet | 89.63 | 81.87 | 92.01 | U-Net | 92.77 | 86.91 | 94.46 | RefineNet | 93.58 | 88.63 | 95.36 | PSPNet | 95.31 | 91.89 | 97.98 | BiSeNet | 95.57 | 92.37 | 98.12 | DeepLabv3+ | 95.96 | 93.26 | 98.69 | Proposed algorithm | 97.18 | 95.03 | 99.22 |
|
Table 1. Performance comparison of different semantic segmentation algorithms for water level scale segmentation
Edge refinement | cSE | sSE | mIoU /% | MPA /% |
---|
| | | 95.96 | 98.69 | | √ | | 96.27 | 98.75 | | | √ | 96.64 | 98.93 | | √ | √ | 96.86 | 99.03 | √ | | | 96.71 | 98.86 | √ | √ | | 96.85 | 98.92 | √ | | √ | 97.02 | 99.14 | √ | √ | √ | 97.18 | 99.22 |
|
Table 2. Experimental results of ablation by introducing channel attention, spatial attention, and edge thinning module
Algorithm | Pixel error | Pixel error rate /% |
---|
SegNet | 13.38 | 3.18 | U-Net | 8.04 | 2.09 | RefineNet | 7.96 | 2.02 | PSPNet | 5.94 | 1.38 | BiSeNet | 6.63 | 1.55 | DeepLabv3+ | 5.48 | 1.22 | Proposed algorithm | 3.73 | 0.76 |
|
Table 3. Horizontal detection effect comparison of semantic segmentation algorithm
Algorithm | Pixel error | Pixel error rate /% |
---|
Algorithm 1 | 51.47 | 11.03 | Algorithm 2 | 14.61 | 3.72 | Proposed algorithm | 3.73 | 0.76 |
|
Table 4. Comparison of horizontal detection effects of different water level measurement algorithms
Moment | Manual reading /m | Algorithmic reading /m | Error /m | Moment | Manual reading /m | Algorithmic reading /m | Error /m |
---|
0:00 | 0.863 | 0.867 | 0.004 | 12:00 | 0.808 | 0.814 | 0.006 | 1:00 | 0.851 | 0.850 | -0.001 | 13:00 | 0.795 | 0.798 | 0.003 | 2:00 | 0.848 | 0.852 | 0.004 | 14:00 | 0.786 | 0.792 | 0.006 | 3:00 | 0.845 | 0.850 | 0.005 | 15:00 | 0.789 | 0.794 | 0.005 | 4:00 | 0.837 | 0.846 | 0.009 | 16:00 | 0.793 | 0.798 | 0.005 | 5:00 | 0.834 | 0.830 | -0.004 | 17:00 | 0.810 | 0.807 | -0.003 | 6:00 | 0.828 | 0.819 | -0.009 | 18:00 | 0.813 | 0.818 | 0.005 | 7:00 | 0.818 | 0.813 | -0.005 | 19:00 | 0.815 | 0.809 | -0.006 | 8:00 | 0.821 | 0.820 | -0.001 | 20:00 | 0.831 | 0.822 | -0.009 | 9:00 | 0.822 | 0.827 | 0.005 | 21:00 | 0.848 | 0.834 | -0.014 | 10:00 | 0.826 | 0.819 | -0.007 | 22:00 | 0.862 | 0.856 | -0.006 | 11:00 | 0.803 | 0.800 | -0.003 | 23:00 | 0.864 | 0.846 | -0.018 |
|
Table 5. Comparison and analysis of water level read by algorithm and manual