Research on the Feature Smoothing Algorithm for Point Cloud Data of Large Complex Surfaces Based on Multichannel Convolutional Neural Network

In reverse engineering, the smooth results of the subtle features of the point cloud data will directly affect the precision of machining quality for large and complex curved workpieces. However, many widely used point cloud filters are not as effective as expected for special surface models. This article proposes the feature smoothing algorithm based on a multichannel convolutional neural network (CNN). Specifically, our approach consists of three stages. First, feature extraction operations are performed on the point cloud data, and the feature point set is marked. Second, the input samples of CNN are constructed according to the feature point set and the nonfeature point set. Third, based on the CNN, a locally smooth surface is fitted near the feature point, and the smooth position of the feature point is predicted. Finally, the proposed algorithm is simulated and analyzed with the actual point cloud data of large spherical crown workpieces. The experimental results show that the proposed algorithm has better feature smoothing accuracy and eliminates the data offset caused by other filtering algorithms in the process of surface feature smoothing.