Portrait Relief Modeling and Animation Making through Convolutional Neural Networks

To overcome the shortcomings of traditional portrait relief design, such as strong professional, low efficiency and single effect, a fast modeling method of portrait relief is proposed. First, a synthetic dataset is generated based on the principle of normal preservation. Then, a neural network is constructed and trained to automatically compress the height field. Finally, after sampling the 3D model from multiple viewing directions, a set of relief models are predicted through the network, then, the relief animation is synthesized. The experimental results on the synthetic portrait relief dataset show that the proposed method can generate portrait relief fast with higher efficiency than that of traditional methods. The modeling result is better than the similar methods in quantitative analysis index PSNR, which has reasonable depth level, and retains the detailed features of human face. The animation provides a novel artistic effect for portrait observation.