Optical Instruments, Volume. 45, Issue 2, 46(2023)
An image semantic segmentation algorithm with a two-branch structure
Figures & Tables(8)
Table 1. Experimental parameter settings
View tableView in ArticleTable 1. Experimental parameter settings
数据集 批处理 大小 初始 学习率 权重衰减 系数 裁剪 大小 优化 函数 Cityscapes 10 0.01 0.0005 1024×1024 SGD CamVid 8 0.001 0.0001 480×360 Adam
Table 2. Ablation experiment results
View tableView in ArticleTable 2. Ablation experiment results
Method mIoU/% Parameters/106 baseline 57.6 11.8 CB 59.4 12.7 CB+DB（sum） 61.5 13.1 CB+DB（FM） 62.5 13.3
Table 3. Segmentation accuracy of different algorithms on the Cityscapes test set
View tableView in ArticleTable 3. Segmentation accuracy of different algorithms on the Cityscapes test set
算法 预训练 主干网络 分辨率 均交并比/% 参数量/106 SegNet Yes VGG16 360×640 56.1 29.5 ENet No No 512×1024 58.3 0.4 SQ Yes SqueezeNet[ 24 ]1024×2048 59.8 — FRRN A No No 256×512 63.0 17.7 DeepLab Yes VGG16 512×1024 63.1 37.3 FCN-8s — VGG16 1024×2048 65.3 134.5 Dilation10 Yes VGG16 1024×2048 67.1 — DeepLabv2 Yes ResNet101 1024×2048 70.4 44.0 RefineNet-LW Yes ResNet101 — 72.1 46.0 RefineNet — ResNet101 1024×2048 73.6 118.0 Ours No ResNet18 1024×2048 67.4 13.3
Table 4. Segmentation accuracy of different algorithms on the CamVid test set
View tableView in ArticleTable 4. Segmentation accuracy of different algorithms on the CamVid test set
Method DeconvNet ENet SegNet FCN-8s BiSeNet 文献[ 26 ]Ours Building — 74.7 88.8 77.8 83.0 — 75.8 Tree — 77.8 87.3 71.0 75.8 — 66.6 Sky — 95.1 92.4 88.7 92.0 — 89.5 Car — 82.4 82.1 76.1 83.7 — 77.0 Sign — 51.0 20.5 32.7 46.5 — 35.1 Road — 95.1 97.2 91.2 94.6 — 93.2 Pedestrian — 67.2 57.1 41.7 58.8 — 39.6 Fence — 51.7 49.3 24.4 53.6 — 26.9 Pole — 35.4 27.5 19.9 31.9 — 17.0 Sidewalk — 86.7 84.4 72.7 81.4 — 78.5 Bicyclist — 34.1 30.7 31.0 54.0 — 44.5 mIoU/% 48.9 51.3 55.6 57.0 68.7 69.1 58.5 Parameters/106 252 0.4 29.5 134.5 49.0 62.0 13.3
Tools
Get Citation
Copy Citation Text
Bing WANG, Qi HU, Yalin BIAN. An image semantic segmentation algorithm with a two-branch structure[J]. Optical Instruments, 2023, 45(2): 46
Paper Information
Category: DESIGN AND RESEARCH
Received: Apr. 18, 2022
Accepted: --
Published Online: Jun. 12, 2023
The Author Email: HU Qi (hq_0519@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20