Bas-relief Design And Optimization Based On Deep Learning

Bas-relief is a kind of artwork that shows the details of three-dimensional objects on a plane.It is a combination of sculpture and painting,which uses compression to process three-dimensional objects.At present,there are three problems with the traditional bas-relief modeling algorithm:First,the generation efficiency of the bas-relief is low.Second,the number of vertices of the 3D model limits the quality of the bas-relief modeling.Finally,the process of the bas-relief stylization algorithm is complicated.In recent years,the rapid development of deep learning technology has achieved breakthrough results in many aspects of image processing.The bas-relief model can be regarded as a height map of 2.5D.Therefore,it is meaningful to design and optimize the process of bas-relief modeling by using deep learning technology to solve the current problems in bas-relief modeling.In order to solve the three problems in the current bas-relief modeling process,this paper has done the following research work:First,traditional methods use the gradient information of the normal graph to construct and solve the corresponding Poisson equation and obtain the final height field.As the resolution of the input normal graph increases,the solution speed will decrease greatly as the scale of the equation increases.In order to solve this problem,we first use the bas-relief modeling method based on normal graph to obtain a bas-relief generated data set,and then construct a fully convolutional neural network whose input is a normal graph and the output is a height graph.Using the neural network to establish the mapping from the normal map to the final height field,it can avoid solving complex Poisson equations.The experimental results show that the algorithm significantly improves the generation speed of the bas-relief model while ensuring the quality of the bas-relief model.Second,due to lack of detail,3D models with low vertex count cannot generate high-quality bas-relief models.In order to solve this problem,we first use the relevant rendering algorithm to generate normal maps of different resolutions.Then we perform super-resolution reconstruction on the low-resolution normal map,and finally generate a bas-relief from the reconstructed normal map.Experimental results show that:compared with the traditional bas-relief modeling algorithm,the algorithm in this paper can effectively improve the quality of bas-relief 3D model of bas-relief modeling.Finally,the traditional bas-relief stylized algorithm extracts the high and low frequency information of the original bas-relief height map,and then re-edits to construct and solve the corresponding Poisson equation to obtain the final height field.The flow of this algorithm is complicated and the production efficiency is low.In order to solve this problem,we apply the style conversion algorithm in deep learning to the field of bas-relief generation.We constructed and trained a convolutional neural network for arbitrary style conversion of normal graphs,and bas-relief modeling of stylized normal graphs to obtain bas-relief models.Experimental results show that the algorithm can generate different patterns of bas-relief models by selecting different stylized normal graphs as input.Compared with traditional algorithms,the process is more streamlined and the production efficiency is higher.