引用本文: 景慧莹, 刘欣怡, 张永军, 任维成, 王雷, 姚永祥, 郭延庆. 低秩矩阵辅助的无人机影像建筑物立面纹理自动修复[J]. 武汉大学学报 ( 信息科学版).
JING Huiying, LIU Xinyi, ZHANG Yongjun, REN Weicheng, WANG Lei, YAO Yongxiang, GUO Yanqing. Low-Rank Matrix Aided Automatic Texture Inpainting of Building Facades from UAV Images[J]. Geomatics and Information Science of Wuhan University.
 Citation: JING Huiying, LIU Xinyi, ZHANG Yongjun, REN Weicheng, WANG Lei, YAO Yongxiang, GUO Yanqing. Low-Rank Matrix Aided Automatic Texture Inpainting of Building Facades from UAV Images[J]. Geomatics and Information Science of Wuhan University.

## Low-Rank Matrix Aided Automatic Texture Inpainting of Building Facades from UAV Images

• 摘要: 无人机数据采集过程中受飞行高度、扫描视场角的限制,使得数据中的建筑物立面纹理存在遮挡问题进而导致数据缺失。然而当前大量待修复数据都是采用人工勾选方式进行修复,人力成本高,时间消耗大。基于此,从建筑物立面纹理低秩性和遮挡区域稀疏性切入,将建筑物立面纹理修复问题化为低秩矩阵补全问题,提出一种低秩矩阵重建辅助的建筑物立面纹理自动修复算法。该算法首先利用鲁棒主成分分析自动完成立面纹理稀疏遮挡区域粗提取;然后,结合遮挡区域的先验条件填充精化得到的稀疏区域作为矩阵补全的约束条件;最后,在矩阵补全模型的基础上,采用非凸秩近似函数截断Schatten-p范数替代核范数,同时又引入总变分算子来解决建筑物立面纹理遮挡区域修复问题。仿真实验和真实数据结果表明:在建筑物低秩纹理修复中,所提算法的分割效果不亚于人工标记,修复效果优于现有算法,能自动提取出待修复区域,实现了人力解放和修复效果的双提升。

Abstract: Objective: Limited by flight height and scanning angle, the building facade texture is inevitably obscured resulting in missing unmanned aerial vehicle (UAV) data. However, in the face of a large amount of data, the current inpainting methods that needs to manually select the target region appears time-consuming and labor-intensive. Therefore, this paper proposes an automatic inpainting algorithm based on low-rank matrix reconstruction for facade texture, starting from the low-rank property of the building and the sparsity of occlusion, and modeling facade texture repairing as solving a low-rank matrix completion problem. Methods: First, a robust principal component analysis algorithm is utilized to automatically complete the coarse extraction of the sparse occlusion of the facade texture; Second, the sparse region obtained by filling and refining with the a priori conditions of the occlusion is used as a constraint for matrix completion; Finally, on the basis of the matrix completion, by adding the truncated Schatten pNorm and the total variation operator to correctly repair the occlusion of the facade. Result: The result show that: Our segmentation is as good as the manual marking, inpainting outperforms many state-of-the-art approaches. Conclusions: Our method can automatically and accurately extract the target region, achieving the double improvement of the manpower liberation and the performance.

/

• 分享
• 用微信扫码二维码

分享至好友和朋友圈