Qianqian Chen, ZhaoWen Li, Shujin Lin
Oct 1, 2022
2022 9th International Conference on Digital Home (ICDH)
Because ordinary people (non-artists) lack professional painting skills, they are often unable to make accurate sketches of visual perception or ideas. In the process of drawing, if there are missing parts of the object and distortion of the strokes, it will affect the further application of the hand-drawn sketch. On the contrary, natural language is the most commonly used and simpler way for people to describe the world. Therefore, we propose an end-to-end network SGBP-AttnGAN to obtain the mapping relationship from end-to-end text to scene level sketches. Aiming at the characteristics of the data modality of sketches, SGBP-AttnGAN mainly includes two parts: sketch multi-scale selection feature fusion network SGNet and sketch multi-mode fusion feature extraction network BPNet. Extensive experiments and comparisons with the original AttnGAN demonstrate that our method is more suitable for the given text description in terms of semantic content and visual perception.