计算机科学与探索2026，Vol.20Issue(5)：1478-1490,13.DOI:10.3778/j.issn.1673-9418.2505047
基于快速采样扩散模型的网络地图端到端生成方法
End-to-End Online Map Generation via Fast-Sampling Diffusion Model
摘要
Abstract
Intelligent online map generation technology has emerged as a prominent research focus in the geographic information field.Current generative model-based mainstream methods for automatic generation fail to make a trade-off between generation speed and quality,which hampers the efficiency of map generation in face of some emergencies like disaster management.To address this challenge,an improved diffusion model-based method is proposed for efficient online map generation.Through extracting the status entropy of the denoising process and designing a feature-level loss function,the proposed method achieves fast and accurate online map generation.Specifically,a discriminator is introduced based on the generative adversarial network to accept the denoising results from the last-time process.The status entropy is calculated as the next time noise,which mitigates the information gap.The residual network-based U-Net network is optimized to achieve smooth outputs of map elements,enhancing the visual quality of the maps.The high-dimensional feature alignment loss function is constructed to constrain the semantic information and boundary of the generated maps.Two public datasets,consisting of remote sensing images of different areas and conditions,are integrated to provide data support.Compared with other mainstream algorithms,the results show that the performance has been improved by 1.2%,2.7%and 18.0%at least in PSNR,SSIM,ACC respectively.The ablation study and converge experiment further validate the outstanding performance on the generation speed and quality.关键词
遥感影像/网络地图生成/扩散模型/状态熵/快速采样Key words
remote sensing imagery/online map generation/diffusion model/status entropy/fast sampling分类
信息技术与安全科学引用本文复制引用
田纪龙,丁肇禛,陈浩,熊伟,邵瑞喆..基于快速采样扩散模型的网络地图端到端生成方法[J].计算机科学与探索,2026,20(5):1478-1490,13.基金项目
国家自然科学基金(42471403,42101435,42101432,62106276).This work was supported by the National Natural Science Foundation of China(42471403,42101435,42101432,62106276). （42471403,42101435,42101432,62106276）
