Abstract: To improve the efficiency of high-precision DEM inpainting process, when the data has large empty irregular areas and significant differences with auxiliary model in scale characteristics and pose, CPU-GPU parallel1 strategy is proposed based on least Z-difference algorithm, which can restore data rapidly without preprocessing,sample training or massive calculation. The method utilizes the minimum distance point searched by Z-coordinate difference to build model, at the same time, the asynchronous flow iteration can be used on the solutions of spatial transformation model parameters. By the simplified algorithm criteria, the computation efficiency is further improved to realize the fast and high precision matching of DEM data from different sources. Meanwhile, based on the pixel step fusion, the vacancy area in the pending DEM can be filled with smooth edge. The experimental results show that the DEM data repaired by this method has high integrity and accuracy, and the processing performs good acceleration ratio of 9.7 compared with the CPU platform, which provides a useful reference for the restoration and analysis of digital elevation models.

Key words: least Z-difference, three-dimensional matching, GPU parallel acceleration, DEM data hole inpainting

摘要： 针对待修复数字高程模型存在较大面积不规则空洞，且与辅助模型的尺度特征、位姿存在明显差异时，需要对数据进行预处理、样本训练或进行大量计算，导致修复效率低下以及产生计算误差等问题，提出了一种基于最小高差（least Z-difference，LZD）匹配和CPU-GPU并行加速的处理方法。利用Z坐标差值搜索最小距离点，将异步流迭代方式应用于空间转换模型参数的求解中，在简化的计算准则基础上进一步提升运行效率，实现对异源DEM空洞数据与基准数据的快速高精度匹配，同时，使用台阶融合算法填补空缺区域并进行边缘平滑。实验结果表明，该方法修复得到的DEM数据完整度、正确性高，处理过程较CPU多线程平台最高能够达到9.7倍的加速比，为数字高程模型的修复与分析研究提供了有益参考。

关键词: 最小高差算法, 三维表面匹配, GPU并行加速, DEM数据空洞修复