Research On Unsupervised Stylized Relief Generation Based On Normals

Bas-relief is an art form between painting and sculpture.It uses a flat or curved surface as a carrier to compress a three-dimensional shape into a limited space.It is widely used in industries such as shape decoration,ceramic utensils,industrial design,commemorative coins,badge production,etc.In the process of bas-relief generation,the normal domain is easier to edit than the depth domain,and has become a current research hotspot.This paper combines unsupervised learning to conduct research on the normal estimation of point cloud models,bas-relief generation and relief stylization.The specific research contents are as follows:(1)An end-to-end normal estimation method suitable for projected point clouds is studied,the network structure of U-Net is improved and the network loss function is optimized.The experimental results show that compared with the existing methods,the algorithm in this paper has higher accuracy,better robustness,and certain advantages in operating efficiency,which can provide a good data basis for subsequent bas-relief generation.(2)An unsupervised relief generation algorithm is studied.Aiming at the problem that the traditional bas-relief modeling algorithm needs to solve the large-scale Poisson equation and the generation efficiency is not high,a neural network is used to complete the mapping from the normal map to the bas-relief height field.In this process,the optimization objective is constructed with the vertical relationship between the normal direction of the vertex and the vector formed by the adjacent points,and the unsupervised training is realized.Experiments have proved that the unsupervised relief generation network in this paper can not only obtain high-quality reliefs with excellent detail features,but also significantly improve the generation speed compared with traditional methods,and there is no need to make "Groudtruth",saving time and effort.(3)Aiming at the problem that the generated relief style is relatively monotonous in the current bas-relief generation field,a normal-based style transfer algorithm is studied.Using the ability of convolutional neural network to extract features,the texture features of the style normal map are transferred to the content normal map,and constraints are added at the same time,so that the geometric structure of the original 3D model is better maintained,and the artifact problem is also improved.Experiments have proved that the relief stylization algorithm in this paper has a good integration of the style texture and the original model as a whole,and can generate a series of basrelief models with different styles,which expands the creation method of bas-relief,and can better meet the creative needs of customers.