测绘通报 ›› 2023, Vol. 0 ›› Issue (12): 153-158.doi: 10.13474/j.cnki.11-2246.2023.0376

• 技术交流 • 上一篇    

实景三维模型纹理的敏感目标自动识别与脱密方法

徐海燕1,2, 郭为人3,4, 李德民1,5, 郝君1,4, 徐刚1,6   

  1. 1. 浙江安防职业技术学院, 浙江 温州 325016;
    2. 温州市地理空间信息技术研究院, 浙江 温州 325016;
    3. 温州市自然资源和规划信息中心, 浙江 温州 325000;
    4. 温州市天空地态势感知应用技术协同创新中心, 浙江 温州 325016;
    5. 温州市自然灾害遥感监测预警重点实验室, 浙江 温州 325016;
    6. 温州市未来城市研究院, 浙江 温州 325016
  • 收稿日期:2023-08-22 发布日期:2024-01-08
  • 通讯作者: 徐刚。E-mail:20096342@zjcst.edu.cn
  • 作者简介:徐海燕(1979-),女,高级工程师,主要从事二三维地理信息技术研究与应用工作。E-mail:20096347@zjcst.edu.cn
  • 基金资助:
    温州市基础性科研项目(S20210017);浙江省自然资源厅2021年度科技项目(2021-34)

Secret target automatic recognition and decryption method for real scene 3D model texture

XU Haiyan1,2, GUO Weiren3,4, LI Demin1,5, HAO Jun1,4, XU Gang1,6   

  1. 1. Zhejiang College of Security Technology, Wenzhou 325016, China;
    2. Wenzhou Institute of Geospatial Information Technology, Wenzhou 325016, China;
    3. Wenzhou Natural Resources and Planning Information Center, Wenzhou 325000, China;
    4. Wenzhou Collaborative Innovation Center for Space-borne, Airborne and Ground Monitoring Situational Awareness Technology, Wenzhou 325016, China;
    5. Wenzhou Key Laboratory of Natural Disaster Remote Sensing Monitoring and Early Warning, Wenzhou 325016, China;
    6. Wenzhou Institute of City Future, Wenzhou 325016, China
  • Received:2023-08-22 Published:2024-01-08

摘要: 实景三维是国家新型基础设施的重要组成部分,在各行业具有广泛的应用。如何在安全的前提下促进三维数据最大程度的共享已成为实景三维应用的迫切需要。针对实景三维纹理中存在涉密敏感目标问题,传统依赖人工检索敏感目标并通过图像编辑工具处理的脱密方法效率低,本文提出一种结合深度学习的纹理影像敏感目标自动识别与脱密方法。首先通过秘密点POI检索包含秘密目标的三维模型及对应的纹理影像;然后基于YOLOv5s网络模型自动识别纹理影像中的敏感目标,并利用GrabCut有效提取目标;最后基于多尺度的Patch match对纹理影像块进行修复。试验表明,本文方法目标识别准确率为95.3%,与人工处理相比全流程用时缩短40%以上,有效提取并去除纹理影像中的敏感目标,实现了快速脱密,促进了实景三维模型数据的安全共享。

关键词: 实景三维, 纹理脱密, 目标识别, 纹理修复

Abstract: Real-scene 3D is an important part of the country's new infrastructure and has a wide range of applications in various industries. How to promote the maximum sharing of 3D data under the premise of safety has become the demand of real-scene 3D applications. Aiming at the problem of secret-related sensitive targets in real-scene 3D textures, the traditional decryption process that relies on manual retrieval of sensitive targets and images processed by editing tools is not efficient. This paper proposes an automatic recognition and decryption method for real scene 3D model texture combined with deep learning. Firstly, search the 3D model and texture image containing the secret target through the secret areas POI; then automatically identify the sensitive target in the texture image based on the YOLOv5s network model, and use GrabCut to effectively extract the target; finally, based on the multi-scale Patch match for the texture image block make repairs. It shows that the target recognition accuracy of the method is 95.3%, which is more than 40% compared with manual processing when the whole process is used. It effectively extracts sensitive targets to achieve fast decryption, and promotes the safe sharing of real-scene 3D model data.

Key words: real scene 3D, texture decryption, target recognition, texture repair

中图分类号: