Research On Terrain Generation Method Based On Deep Learning

As an important factor in virtual scenes,virtual terrain is widely used in many fields,such as virtual military exercise,virtual reality and 3D simulation.How to generate user-controlled terrain quickly and efficiently has always been a challenging topic in the field of graphics.In recent years,the rapid development of deep learning technology has shown the effects that traditional algorithms are difficult to match in many fields,such as landscape painting migration,virtual portrait synthesis and so on.How to generate user-controlled and highly realistic terrain quickly and effectively based on deep learning technology has become a new research direction.In the generation model based on deep learning,GAN model has gradually become the preferred model for generating tasks with its excellent generation effect.Among the various branches of GAN,CGAN provides constraints on the generated results and increases the control over the generated images.Pix2 pix is a general image translation network proposed on the basis of CGAN.Pix2 pix can be directly applied to the translation task from valley line and ridge line map to topographic map,and some results can be achieved.However,the training of the model is more difficult,and the terrain generation effect is not ideal when the input ridge line is relatively sparse,and there will be obvious repeated grid traces that do not conform to the natural law in the topographic map.In order to solve the above problems,this thesis designed the Topo GAN network suitable for terrain generation based on the structure of the original pix2 pix network and the characteristics of DEM topographic maps.Topo GAN network can automatically generate the terrain map that meets the user's expectation according to the description of terrain features such as ridge line and valley line provided by the user,and can effectively solve the problem of grid trace.The main work and contributions of this thesis are summarized as follows:1.For the terrain generation task,this paper introduces the pix2 pix network based on CGAN,and analyzes the basic principle and network structure of the model.And we finds that the model of the network is complex,and there will be obvious repeated grid traces in the output terrain when the sparse ridge line condition is input.2.Combined with the characteristics of DEM topographic map and the problems ofpix2 pix network in terrain generation,this thesis proposes a Topo GAN network which is more suitable for terrain generation based on pix2 pix network.Topo GAN reduces the training complexity of the network by changing the information interaction mode of the cross layer channels in the generator,and avoids the problem of repeated grid traces in the generated topographic map by changing the expansion mode of the characteristic map in the decoding part of the generator.3.In this thesis,several test samples are used to test the effectiveness of Topo GAN in terrain generation.This thesis also compares the terrain generation effect of pix2 pix and Topo GAN under sparse ridge line,and compares the training complexity of them.Then,the data set of Qinling geomorphic style is made,and the terrain generation effect of Topo GAN under the data set is tested.The experimental results show that the Topo GAN network proposed in this thesis can generate realistic three-dimensional terrain according to the valley line and ridge line.Compared with pix2 pix network,Topo GAN reduces the training complexity,solves the problem of repeated grid traces in the terrain generated by pix2 pix network,and produces better terrain effect when the ridge line is sparse.Topo GAN can effectively generate Qinling style terrain,which verifies the ability of model learning and restoring specific landform.