A dynamic displacement map based on deep Q-network to assist the rendering of stylised 3D models

In the process of rendering with NPR, there will often be a problem of reduced perception of the NPR effect caused by the fixed spatial structure of the static mesh. Therefore, this research first establishes a convolutional neural network to conduct supervised training on numerous hand-drawn target stylised images, and perform continuous angle recognition through stylised images then output its angle vector. Secondly, the displacement map is generated in real-time by inputting the observation angle through the fully connected neural network model. After that, the real-time displacement map is sampled to dynamically generate the deformation of the model. In the end, the goal of breaking the model's sense of space and enhancing the NPR rendering effect can be achieved. Besides, the experimental results of this study verify the effectiveness of the method.