Laser & Optoelectronics Progress, Volume. 61, Issue 8, 0837004(2024)
A Pointer Meter Bilateral Image Segmentation Network Integrating Spatial Details and Semantic Features
Figures & Tables(9)
Fig. 2. Comparison of the detail branch convolution block before and after adjustment. (a) Detail branch convolution block of BiSeNet V2; (b) ConvNeXt convolution block
Fig. 3. Comparison of semantic branch structures. (a) BiSeNet V2 primitive semantic branch; (b) improved semantic branch
Fig. 4. Comparison of the structure of bilateral-guided aggregation layer and bilateral-guided splicing aggregation layer. (a) Bilateral-guided aggregation layer of BiSeNet V2; (b) bilateral-guided splicing aggregation layer
Fig. 5. Visualization of instrument segmentation results. (a) BiSeNet V2; (b) improved detail branch; (c) improved semantic branch; (d) improved aggregation layer; (e) ground truth
Fig. 6. Comparison of the results of different backbone networks in instrument image segmentation. (a) Ground truth; (b) ViT; (c) BEiT; (d) Swin-T; (e) Twins; (f) PoolFormer; (g) BiUnet (ours); (h) MobileNetV3; (i) Fast-SCNN; (j) CGNet; (k) BiSeNet V2; (l) ConvNeXt
Table 1. Comparison of detail branch parameters of original and improved BiSeNet V2
View tableTable 1. Comparison of detail branch parameters of original and improved BiSeNet V2
Stage BiSeNet V2 detail branch Improved detail branch Output size Opr p k C s r Opr p k C s r Input 512×512 1 Conv2d 1 3 64 2 1 Conv2d 1 2 384 2 1 256×256 Conv2d 1 3 64 1 1 ConvNeXt 3 7 384 1 1 2 Conv2d 1 3 64 2 1 Conv2d 1 2 384 2 1 128×128 Conv2d 1 3 64 1 2 ConvNeXt 3 7 192 1 3 3 Conv2d 1 3 128 2 1 Conv2d 1 2 192 2 1 64×64 Conv2d 1 3 128 1 2 ConvNeXt 3 7 128 1 3
Table 2. Ablation experimental results
View tableTable 2. Ablation experimental results
Detail branch Semantic branch Aggregation layer IoU /% mIoU /% #Param. /106 Background Line Pointer_mid Pointer Baseline Baseline BGA 98.54 58.15 85.91 77.49 80.02 3.34 Improved Baseline BGA 98.88 65.68 89.35 85.93 84.96 7.19 Baseline Improved BGA 98.96 67.98 89.55 86.01 85.62 3.26 Improved Improved BGA 98.94 68.25 89.79 84.88 85.47 5.76 Improved Improved BGCA 99.29 78.02 89.67 87.67 88.66 4.93
Table 3. Performance of BiUet algorithm and other algorithms on the instrumentation dataset
View tableTable 3. Performance of BiUet algorithm and other algorithms on the instrumentation dataset
Backbone IoU /% mIoU /% mAcc /% aAcc /% #Param. /106 Flops /109 Background Line Pointer_mid Pointer PoolFormer［ 16 ］97.11 16.19 81.63 44.11 59.76 65.60 97.13 12.59 9.82 Twins［ 19 ］97.21 15.06 83.14 46.00 60.35 66.39 97.23 25.95 16.44 Swin-T［ 18 ］97.20 15.28 83.46 46.77 60.68 66.44 97.22 31.06 25.59 BEiT［ 17 ］97.74 29.47 88.72 73.51 72.36 78.05 97.77 23.96 18.46 ViT［ 15 ］97.75 30.30 88.85 74.90 72.95 78.87 97.77 23.76 16.65 ConvNeXt-T［ 13 ］97.26 17.14 84.68 48.97 62.01 67.91 97.28 34.92 25.11 BiSeNet V2［ 12 ］98.54 58.15 85.91 77.49 80.02 87.60 98.57 3.34 12.31 CGNet［ 20 ］98.74 61.25 86.01 81.98 81.99 87.93 98.76 16.64 25.95 Fast-SCNN［ 21 ］98.82 64.05 87.72 84.43 83.75 89.49 98.84 1.95 3.19 MobileNetV3［ 22 ］98.88 65.39 87.68 84.81 84.19 89.77 98.90 11.83 42.26 BiUnet（ours） 99.29 78.02 89.67 87.67 88.66 93.72 99.30 4.93 20.82
Tools
Get Citation
Copy Citation Text
Yaohui Zhu, Zhigang Wu, Min Chen. A Pointer Meter Bilateral Image Segmentation Network Integrating Spatial Details and Semantic Features[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0837004
Paper Information
Category: Digital Image Processing
Received: Jun. 5, 2023
Accepted: Jul. 24, 2023
Published Online: Mar. 13, 2024
The Author Email: Wu Zhigang (wzgang2017@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20