关键词: 逆向工程, 三角形网格, 神经网络, 逼近误差, 边缘误差, 散乱点云

Abstract: An approach based on the self-organizing feature map (SOFM) neural network has
been developed to reconstruct Delaunay triangle mesh for the unorganized measured point cloud.
However the approach suffers from approximation and boundary problems. A triangle mesh model
with high approximation precision is proposed in order to reduce the approximation error and
boundary error. First all the neurons of the mesh model are trained directly over the unorganized
point cloud. Next the neuron location weights of the mesh model are adjusted along the normal
vectors of the mesh vertices. Last only the boundary neurons of the mesh model undergo training
by the boundary points of the measured point cloud. As a result of applying the proposed mesh
model, the boundary error is greatly reduced and the mesh is drawn toward the sampled object
with higher precision comparing with the original SOFM training algorithm. The feasibility of the
developed mesh model is demonstrated on two examples.

Key words: reverse engineering, triangle mesh, neural network, approximation error;
boundary error, unorganized point cloud