三维残缺数据的神经网络修补方法

3D Incomplete Data Repairing Algorithm Based on Neural Network

针对常用的三坐标测量设备在对物体表面采样时造成的局部区域数据残缺问题,提出了一种三维残缺数据的多层感知器神经网络修补方法。该方法首先用矩形框在残缺数据的边界附近获取样本点集,并以其最小二乘拟合平面为基础建立局部坐标系;然后,在此局部坐标系下,将训练后的MLP网络仿真曲面用于残缺区域数据点重采;最后,将重采点集经坐标反变换后去替代原始点云数据中的样本点集。仿真和真实残缺数据的修补实验结果表明,此方法具有较高的修补效率和精度,可获得满意的修补效果。

Due to surface reflection property, occlusion and accessibility limitation, certain areas of the object were usually not sampled, leading to holes and undesirable artifacts in the resulting models. To improve the quality of modeling, a new repairing algorithm for 3D incomplete data was proposed. First, the sample points were acquired around the incomplete data boundary by using a rectangle frame, and a local coordinates system based on the least square-fitting plane of the samples was established, Then, in this coordinates system, the new points were resampled over the incomplete region by the trained multi-layer perceptrons （MLP） network, Finally, the resampled points were used to displace the sample points in cloud data after these points were transformed from the local coordinates system to the original coordinates system. Experimental results show that the algorithm has higher repairing efficiency and precision.