A sketch semantic segmentation method based on conditional generative adversarial and mean teacher models

A sketch semantic segmentation method based on conditional generative adversarial and mean teacher models

연구 분야: Strategies

학회: Signal, Image and Video Processing

Most existing sketch semantic segmentation methods rely on stroke sequence information to achieve high performance, which makes them unable to handle static sketches without stroke sequence information. In this paper, we propose a sketch semantic segmentation method based on conditional generative adversarial and mean teacher models to avoid dependence on stroke sequence information. Specifically, a relational graph convolutional neural network is constructed, which utilizes several nearest neighbor relationships and a pseudo stroke continuity relationship to understand the semantics of sketches. Based on the generative adversarial idea, we design a conditional generative adversarial model to use sketches without semantic annotations for improving the semantic segmentation performance. Subsequently, to limit the incorrect semantic types, we implement a semantic classification network by building a mean teacher model, aiming at using sketches without semantic annotations to enhance the semantic classification performance. Finally, the semantic classification results are combined with the semantic segmentation outputs of the conditional generative adversarial model to obtain the final results. The extensive experiments on SPG and SketchSeg-150K demonstrate that the proposed method achieves satisfactory performance without stroke sequence information.