Mesh model registration via deep mapping by local projection for optical scanning based reverse engineering

Objective: For mesh model obtained by optical scanning in reverse engineering, sub models registration to be the complete model is a key step. However, currently widely used registration methods may need well initial position condition. Moreover, in some common adverse conditions (e.g. with large angle differences, with incomplete common parts), they cannot achieve registration for sub models. We solve these issues from a direct matching viewpoint and present a robust and straightforward learning based registration method for 3D mesh models with any initial position conditions. Methods: First, projection is utilized to collect local geometrical information of all the vertices within the 3D mesh model; Second, the information of each vertex is stored in a matrix; Third, the Convolutional Auto-Encoder is utilized to reduce the matrices and extract main features of the geometrical information; Fourth, mesh vertices are matched by comparing the similarity of corresponding feature matrices; Finally, the RANSAC algorithm is utilized to register the 3D mesh models based on the mapping vertices.Result: Comparing with other classic methods, it not only can achieve registration with any initial angle dif-ference, but also can complete registration with inclusion relation and common parts.