| 摘 要: 传统点云修补方法对于特征不明显的点云修补效果不佳,为此提出了一种基于深度学习的点云修补算法,算法框架包括自动编码器、对抗生成网络及强化学习模块。该算法首先在潜在空间表征上训练对抗生成网络,然后引入自动编码器网络,自动编码器网络中的编码器可以将输入点云压缩成潜在空间表征,最后引入强化学习模块,为对抗生成网络的生成器选择合适的向量合成完整点云的潜在空间表征。实验表明,该算法可以修补特征不明显的点云,并能在毫秒级别内较为有效地完成点云缺失区域修补。 |
| 关键词: 点云修补;深度学习;对抗生成网络 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 黄河交通学院智能汽车计算机视觉工程研究中心项目(22KYGJ01). |
|
| Real-Time Point Cloud Repair Algorithm Based on Deep Learning |
|
QIAO Yaoyao, ZHANG Lanlan
|
(College of Intelligent Engineering, Huanghe Jiaotong University, Jiaozuo 454950, China)
1157764920@qq.com; 2089208988@qq.com
|
| Abstract: Traditional point cloud repair methods have poor performance in repairing point clouds with unclear features. Therefore, this paper proposes a point cloud repair algorithm based on deep learning, which includes an automatic encoder, adversarial generation network, and reinforcement learning module. This algorithm first trains the Generative Adversarial Network (GAN) on potential spatial representations. Then, an automatic encoder network is introduced, the encoder of which can compress the input point cloud into potential spatial representations. Finally, a reinforcement learning module is introduced to select suitable vectors for the generator of the GAN to synthesize the potential spatial representations of the complete point cloud. Experiments have shown that the proposed algorithm can repair point clouds with unclear features and effectively repair missing areas of point clouds within milliseconds. |
| Keywords: point cloud repair; deep learning; GAN |
