Surface reconstruction of disordered point cloud based on adaptive learning neural network

Currently existing methods to reconstruct B-spline surfaces through neural network do not consider the endpoint and tangential constraints, making it unsuitable for surface reconstruction with such constraints. In view of that, this paper proposes a surface reconstruction method of disordered point cloud based on adaptive learning neural network. Under the proposed neural network structure, we take the point cloud parameters as input layer, the physical coordinates corresponding the parameters as output layer, and the control points to be solved as weight coefficients. The constraints are introduced into the loss function by Lagrange multiplier method, and the adaptive learning rate is designed to accelerate the learning process. In addition, a parametric process is designed. First, the disordered point cloud is mapped to a two-dimensional (2D) surface. Then, the point cloud parameterization is completed by constructing a parametric surface accurately matching the point cloud mapped to the 2D surface. Finally, we compare the traditional method and the proposed method without any constraints imposed initially, and the results show superior performance of the new method in terms of reconstruction accuracy. Adding tangential constraints, experimental results show the overall reconstruction accuracy can still be maintained under the proposed method.