| 摘要: |
| 高分辨率无线电地图的构建面临全局依赖难以建模与局部结构难以保持的双重困难挑战。为此，提出了一种融合双重注意力多尺度特征与Transformer的条件生成对抗网络DAMF-TransCGAN。首先，引入改进的深度高分辨率模块以增强信号强度空间分布的局部纹理重构能力；其次，设计双重注意力多尺度特征融合模块，提升多尺度信号强度分布的局部特征提取能力；然后，结合Transformer架构捕捉稀疏观测点间的远距离全局依赖关系；最后，设计结构保持损失函数以增强遮挡区域的重建一致性，适应不同建筑信息的可获取性。在两类经典场景（已知与未知建筑物分布）下展开实验，结果显示，DAMF-TransCGAN在低采样率下均显著优于现有最优模型(RadioUNet和RME-GAN)，尤其在未知建筑物信息场景中，其归一化均方误差和根均方误差相较于RadioUNet分别降低了17.46%和5.19%。 |
| 关键词: 高分辨率无线电地图 条件生成对抗网络 Transformer |
| DOI：10.20079/j.issn.1001-893x.250709001 |
|
| 基金项目:国家自然科学基金资助项目（62371082）；湖南省自然科学基金项目（2024JJ8024） |
|
| DAMF-TransCGAN Generative Model-aided High-precision Radio Map Construction Method |
| DENG Xueshuang,ZHANG Min,YANG Yuyou,SHEN Bin |
| (1.School of Communications and Information Engineering，Chongqing University of Posts and Telecommunications，Chongqing 400065，China;2.College of Information and Communication Engineering，Hunan Post and Telecommunication College，Changsha 410015，China) |
| Abstract: |
| In complex scenarios with sparse sampling and building occlusions，constructing high-resolution radio maps is challenged by two issues：the difficulty of modeling long-range global dependencies and preserving local structural details，both of which are further exacerbated by the lack of prior information such as transmitter locations and building layouts.To address these challenges，a Dual-Attention Multi-Scale Feature Fusion and Transformer-based Conditional Generative Adversarial Network (DAMF-TransCGAN) is proposed.Specifically，an enhanced deep high-resolution module is employed to improve the reconstruction of local textures in signal strength distributions.A dual-attention multi-scale fusion module is further designed to strengthen local feature extraction across different scales.In addition，a Transformer architecture is integrated to capture long-range dependencies among sparse observations.A structure-preserving loss function is introduced to improve reconstruction consistency in occluded regions and adapt to varying levels of building information availability.Extensive experiments under two representative scenarios (with known and unknown building layouts) demonstrate that DAMF-TransCGAN achieves significant improvements over the state-of-the-art models，RadioUNet and RME-GAN，at low sampling rates.In particular，compared with the best-performing baseline，RadioUNet，the proposed model reduces the normalized mean square error and root mean square error by 17.46% and 5.19%，respectively，under scenarios with unknown building layouts. |
| Key words: high-accuracy radio map conditional generative adversarial networks Transformer dual-attention multi-scale feature fusion |
《电讯技术》编辑部 