Chinese Journal of Liquid Crystals and Displays, Volume. 40, Issue 9, 1333(2025)
Lightweight building extraction network integrating wavelet transform and global awareness
Fig. 4. Global context-aware module. （a） GCAM module； （b） CCSA module；（c） CSSA module.
Fig. 5. Building extraction maps for different models on WHU dataset
Fig. 6. Building extraction maps for different models on Inria dataset
Table 1. Network model deployment
View tableView in ArticleTable 1. Network model deployment
项目 具体细节 数据构成 WHU：4 736张训练图像，1 036张验证图像，2 416张测试图像；Inria：9 294张训练图像，2 325张测试图像 数据预处理 WHU：标准化处理，分割为512像素×512像素；Inria：非重叠裁剪为512像素×512像素，滤除无建筑物样本 实验平台 PyTorch框架， NVIDIA GeForce RTX 4060 GPU （8 GB显存） 优化器 Adam优化器，初始学习率0.001，多阶段衰减 损失函数 边界加权二进制交叉熵损失（权重0.7）+骰子损失（权重0.3） 训练设置 批大小：4（训练）、2（验证），200训练周期 数据増强 随机旋转、水平/垂直翻转、尺度缩放等空间变换 评估指标 Precision、Recall、F1-score、IoU、FLOPs、Params、FPS
Table 2. Accuracy indicators of experiment results from different network models on WHU dataset
View tableView in ArticleTable 2. Accuracy indicators of experiment results from different network models on WHU dataset
Methods Precision/% Recall/% F1/% IoU/% Params/M FLOPs/G FPS Swin Transformer 92.83 93.79 93.31 87.45 27.18 30.88 22.71 BuildFormer 93.80 94.32 94.06 88.79 40.52 116.36 15.28 SDSCUNet 93.92 94.45 94.19 89.01 21.32 29.81 24.97 EasyNet 93.97 94.67 94.32 89.25 3.90 17.30 28.67 HDNet 94.61 93.96 94.28 89.18 13.89 201.37 9.89 CaSaFormerNet 93.86 92.15 93.00 86.91 4.44 1.99 29.93 LWGNet 94.20 95.37 94.78 90.08 3.09 4.93 30.24
Table 3. Accuracy indicators of experiment results from different network models on Inria dataset
View tableView in ArticleTable 3. Accuracy indicators of experiment results from different network models on Inria dataset
Methods Precision/% Recall/% F1/% IoU/% Params/M FLOPs/G FPS Swin Transformer 83.68 88.69 86.11 75.61 27.18 30.88 22.71 BuildFormer 85.79 88.38 87.07 77.09 40.52 116.36 15.28 SDSCUNet 86.70 88.09 87.39 77.60 21.32 29.81 24.97 EasyNet 87.24 87.54 87.39 77.61 3.90 17.30 28.67 HDNet 87.61 86.64 87.03 77.04 13.89 201.37 9.89 CaSaFormerNet 87.51 85.48 86.48 76.19 4.44 1.99 29.93 LWGNet 87.50 89.26 88.37 79.16 3.09 4.93 30.24
Table 4. Quantitative evaluation results of ablation experiments on the WHU dataset
View tableView in ArticleTable 4. Quantitative evaluation results of ablation experiments on the WHU dataset
Methods Precision/% Recall/% F1/% IoU/% Params/M FLOPs/G FPS baseline 93.08 93.89 93.48 87.76 6.60 13.50 24.18 baseline+EM 93.95 94.07 94.01 88.07 6.60 13.53 23.56 baseline+GM 92.44 95.35 93.87 88.45 6.43 12.93 25.60 baseline+SSD 93.11 95.37 94.23 89.09 3.26 5.47 29.84 baseline+EM+GM 92.76 95.50 94.11 88.88 6.43 12.96 25.16 baseline+EM+SSD 94.04 94.66 94.35 89.30 3.26 5.50 29.22 baseline+GM+SSD 93.68 95.27 94.47 89.51 3.09 4.90 31.46 baseline+EM+SSD+CCSA 94.81 94.26 94.53 89.63 3.09 4.79 29.56 baseline+EM+SSD+CSSA 94.01 94.98 94.50 89.57 3.09 4.62 29.77 baseline+GM+EM+SSD（ G=2） 94.20 95.37 94.78 90.08 3.09 4.93 30.24 baseline+GM+EM+SSD（ G=1） 94.49 95.14 94.81 90.13 5.07 6.75 29.27
Get Citation
Copy Citation Text
Wen SHAO, Pan SHAO, Baogui SONG, Biao XIONG. Lightweight building extraction network integrating wavelet transform and global awareness[J]. Chinese Journal of Liquid Crystals and Displays, 2025, 40(9): 1333
Category:
Received: May. 22, 2025
Accepted: --
Published Online: Sep. 25, 2025
The Author Email: Pan SHAO (panshao@whu.edu.cn)
