基于PCPNet改进的深度学习点云去噪方法研究

Research on Improved Deep Learning Point Cloud Denoising Method Based on PCPNet

针对三维设备获取的点云数据存在大量噪声和数据处理中局部细节部分丢失等问题,提出一种基于PCPNet改进的深度学习网络来有效去噪。该网络利用多尺度特征聚合模块自适应的聚合了不同尺度点云块的局部细节特征和整体特征,并用LSTM来聚合不同尺度的点云块特征,这样能够更好地保留局部特征,估算更为精确的局部细节法线。多次试验表明,该方法相较于传统方法和PCPNet等深度学习方法在性能上更好,对不同噪声具有一定的鲁棒性,同时又能够对局部边缘信息进行有效保留。

Aiming at the problems of a lot of noise and loss of local details in the point cloud data acquired by 3D equipment,an improved deep learning network based on PCPNet is proposed to effectively denoise in this paper.The network uses the multi-scale feature aggregation module to adaptively aggregate the local detail features and overall features of point cloud blocks of different scales,and uses LSTM to aggregate the features of point cloud blocks of different scales,which can better retain local features and estimate more accurate local detail normals.Multiple experiments show that this method has better performance than traditional methods and deep learning methods such as PCPNet,and has certain robustness to different noises while effectively retaining local edge information.