Technology Study Of Terrain Synthesis Based On Learning Strategy

Technology of terrain synthesis is widely used in film and television creation,game development and military simulation,and its research has certain practical value.The traditional terrain synthesis methods has low controllability.For example,procedural methods and physical erosion based methods mainly aim to satisfy the demand of realistic,but they fail to achieve the realistic effect consistent with the real terrain.Meanwhile,these methods rarely consider the demand of user customization.To overcome the current defects and bottlenecks of traditional methods,this paper explores the terrain synthesis technology based on machine learning strategy,and uses Digital Elevation Model(DEM)as a sample of synthetic terrain to study the key technology of realistic terrain synthesis under the control of user hand sketches.Firstly,this dissertation studys the terrain synthesis based on salient features.Based on the DEM examples,we extarct the salient features by the defined local statistical entropy to build the dataset,and the Radial Basis Function(RBF)neural network is also constructed.After that,user sketches are used to control and synthesize the results with specific user-customized features,also consistent with real terrain.In this section,we propose an effective strategy for user-customized terrain synthesis,and the synthetic results can maintain the salient features of the sample terrain.Secondly,a multi-scale detail fusion terrain synthesis strategy based on deep learning is proposed,which is composed of salient feature sub-network and multi-scale detail retention sub-network.Using the pre-extracted salient features as tag data,the salient feature sub-network is pre-trained,and the multi-scale detail sub-network is further combined to realize the terrain synthesis result with detailed features.The research defines a multi-scale loss function.User sketch is used to control the saliency feature distribution to ensure network's performance,so that the synthetic terrain can meet demand of realistic while satisfying the demand of user customization.In addition,this dissertation explores a multi-feature terrain synthesis strategy based on conditional variational auto-encoder.Generative Adversarial Network(GAN)and the conditional variational self-encoder are combining to construct a conditional variational autoencoder and learning model with multi-feature constraints.To ensure that the synthetic terrain can meet requirements of user customization,the conditional variational self-encoder is constrained by ridgelines,river networks,and geometric feature points of the sample terrain,which can get the results with the consistent features with the sample terrain.Conclusively,this dissertation proposes a new solution to the terrain synthesis problem of user hand sketch control by exploring the key technology of terrain synthesis based on learning mechanism,which provides a reference for the research of personalized terrain customization problem and brings it to the application with certain practical value.