Generating face images from fine-grained sketches based on GAN with global-local joint discriminator

• Published in: 2022 IEEE 2nd International Conference on Software Engineering and Artificial Intelligence (SEAI) pp. 50 - 54
• Main Authors: Gao, Huachao, Mao, Wei, Lin, Yongping
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.06.2022
• Subjects: Analytical models; Conferences; fined-grained sketches; Generative adversarial networks; global-local joint discriminator; Image edge detection; Image synthesis; image-to-image translation; self-attention; Task analysis; Visual effects
• DOI: 10.1109/SEAI55746.2022.9832103

This paper explores the face image generating with clear details from fine-grained sketches. Edge maps are usually used as sketches in face image generation tasks. However, there are some problems such as discontinuous lines and a lack of detailed information in the edge map, so the generated face image is not clear enough and lacks details. To address this problem, a face sketch dataset with rich details is made. The discriminator combines a global discriminator and a local discriminator, which ensures that the generated face image has a complete face structure and generates clearer face details. A self-attention mechanism is employed in the generative model to establish long-term dependencies on partial feature information. The model is quantitatively evaluated by using the valuation criteria IS, FID, and KID. Among them, the evaluation scores of the model on IS, FID, and KID are 1.956 ± 0.053, 17.941 ± 0.970, and 0.009 ± 0.001, respectively. The evaluation results show that the model achieves good performance. Furthermore, the visual effects of the model are analyzed by comparing with the pix2pixHD and CycleGAN models on the generated images. The final results show that our model outperforms the other two models and can generate high-quality face images with sharp details.