Fig. 1. Recovery process

Fig. 2. Relationship between voxel discretization degree and error

Fig. 3. Network structure that outputs 3D voxel models after 2D images are input into the deep learning network

Fig. 4. Registration fusion process between image point cloud and laser point cloud

Fig. 5. Facial missing laser point cloud. (a) Newton statue; (b) Madame Curie statue

Fig. 6. Facial image point cloud generated by the pre-training network compared with the facial image point cloud generated by 3DDFA-V2 network and facial laser point cloud.(a) (f) Point cloud images; (b) (g) point cloud front; (c) (h) point cloud right; (d) (i) 3DDFA-V2 point cloud right; (e) (j) laser point cloud

Fig. 7. Effect comparison of the recovery model obtained by proposed method.(a) (d) Simulation of facial damage model; (b) (e) facial restoration model; (c) (f) three-dimensional model before facial damage

Fig. 8. Effect comparison between the proposed method and the symmetry restoration model.(a) (d) Restoration model of the proposed method; (b) (e) restoration model front based on symmetry; (c) (f) restoration model side based on symmetry

Fig. 9. Visualization results of distance between points with the same name.(a) (c) Visualization results of homonymous point distance between 3D point cloud and laser point cloud generated from single statue face image; (b) (d) visualization results of homonymous point distance between statue face point cloud and laser point cloud after symmetry repair

Table 1. Average distance of homonymous point