关键词: 服装设计, 手绘草图, 深度学习, 图像生成, 生成式对抗网络, 注意力机制

Abstract:

Objective Garment image synthesis is an important part of the garment design and manufacturing process, which uses artificial intelligence technology to automatically generate realistic garment images. Garment design relies heavily on the subjective will of the designer, which often needs to be manually achieved by designers. However, this process is time-consuming and quite inefficient. In the context of artificial intelligence, garment image synthesis can significantly improve efficiency by automatically generating garment images. In addition, it has a wide range of applications in virtual try-on, fashion image manipulation and fashion presentation. Therefore, garment image synthesis has received a lot of attention.
Method The garment sketch was guided to automatically generate the corresponding garment image by entering the garment attributes. A garment image generation method based on hand-drawn sketches was proposed, namely AGGAN. Generative adversarial networks with attention mechanism was adopted to learn garment sketches and garment image data to obtain AdaIN parameters after One-hot encoding of garment attributes through the attribute incorporation module, which are incorporated into the model, and the model was trained to learn the correspondence between garment images and their visual attributes, thus can generate corresponding garment images under the guidance of given garment attributes.
Results AGGAN was qualitatively compared with some existing image generation methods (Fig.2). By comparing with all baselines, the AGGAN proposed not only generates garment images with multiple colors, but also generates images closer to the real situation in terms of visual effects. In addition, IS (inception score), FID (fréchet inception distance), and MOS (mean opinion score) was useds for further quantitatively evaluating the model. The IS value of the garment images generated by the method prosposed is 1.253 (Tab.1), which is 13.8% higher than CycleGAN (cycle-consistent generative adversarial networks ) value, and higher than the values from using other methods. The FID value is 139.634, which is 26.2% lower than CycleGAN, and lower than other methods. In addition to the above two evaluation methods, MOS was adopted to evaluate the quality of garment images generated by each method, the MOS score obtained by the method prosposed is 4.352, which is higher than other image generation methods. In order to control the generation of garment images more flexibly, experiments was conducted on attribute-guided garment image synthesis. The garment sketch was controlled by garment attributes to synthesize the corresponding garment image, and the generated garment image has obvious changes in the sleeve length part, which does not seem to be particularly incongruous (Fig.3). The effect of the color attribute on the generated garment images was also explored. Several common color attributes was chosen in the experiments, and it can be seen that AGGAN can generated almost any corresponding color and high-fidelity garment images under the control of color attributes (Fig.4). Texture is also the most intuitive and the main visual feature of the garment image, and several texture attributes was selected in the experiments (Fig.5). From the figure it can be seen that the generation results are more obvious, basically the required texture can be generated, although further improvement is necessary in terms of realism.
Conclusion The research constructed a garment image generation model based on hand-drawn sketches through the attribute incorporation module, attention mechanism and CycleGAN. Combining the advantages of generative adversarial networks and conditional image generation methods, it took garment attributes as conditions to improve the controllability of the garment image generation process, which helps garment designers to achieve automated garment image synthesis. After a series of experiments, the feasibility and effectiveness of the text method were proved. The method proposed provides new ideas for computer-aided garment design. Some improvements should to be made, for example, the generated garment images cannot generate texture attributes effectively, and there are fewer garment attributes studied.

Key words: fashion design, hand-drawn sketch, deep learning, image generation, generative adversarial network, attention mechanism