【摘要】 针对目前基于深度学习的地形生成研究中仅单独考虑语义图或草图的局限,提出了一种结合地形种类分割语义图和地形特征线(点)草图进行联合生成的多类型地形生成模型。模型基于Pix2Pix框架,使用条件生成对抗网络的思想,在编码器阶段对两种输入的隐层特征进行融合,在生成器中引入带约束的类自适应归一化方法,在损失函数中额外引入VGG提取的纹理损失,生成能够同时兼顾语义图和草图的地形高度图,实现按照语义图划分区域生成不同类型地形且满足草图约束的效果;针对特征草图有效信息占比较少问题,提出了一种特征位置编码方法,提高模型的生成效果;同时制作可用于本实验的数据集对模型进行训练,通过对比实验、消融实验充分验证了所提出方法在地形生成任务中的有效性和优越性。更多还原

【Abstract】 In response to the limitations of current terrain generation research based on deep learning, which only considers either semantic maps or sketches individually, a multi-type terrain generation model is proposed that combines terrain type segmentation semantic maps and terrain feature line(point)sketches for joint generation. The model is built on the Pix2Pix framework and employs the idea of conditional generative adversarial networks. During the encoder phase, the hidden layer features of the two inputs are fused. The generator introduces a constraint-based classadaptive normalization method, and an additional texture loss extracted by VGG is incorporated into the loss function.This enables the generation of terrain height maps that simultaneously consider both semantic maps and sketches, achieving the effect of generating different types of terrain according to the semantic map’s region division while satisfying the constraints of the sketches. To address the issue of the low proportion of effective information in feature sketches, a feature position encoding method is proposed to enhance the model’s generation performance. A dataset suitable for this experiment was also created to train the model. Through comparative experiments and ablation studies,the effectiveness and superiority of the proposed method in terrain generation tasks have been thoroughly validated.更多还原