Automatic Coloring of Anime Line Art Based on Improved CycleGAN

Generative Adversarial Networks as a type of deep learning model, can automatically learn the color and content information in images by learning the distribution of real images, thereby generating realistic color images. In animation production, traditional manual coloring is labor-intensive and ch...

Full description

Saved in:
Bibliographic Details
Published in2024 7th International Conference on Computer Information Science and Application Technology (CISAT) pp. 454 - 457
Main Authors Zheng, Liguo, Huati, Muheyati, Feng, Changbao, Qu, Feng
Format Conference Proceeding
LanguageEnglish
Published IEEE 12.07.2024
Subjects
Accuracy
Analytical models
coloring
Computational modeling
deep learning
GAN
Image color analysis
Line Art
Manuals
Measurement
Production
Transformers
Visual effects
Online AccessGet full text

Cover

More Information
Summary:Generative Adversarial Networks as a type of deep learning model, can automatically learn the color and content information in images by learning the distribution of real images, thereby generating realistic color images. In animation production, traditional manual coloring is labor-intensive and challenging to ensure consistency. GAN-based automatic coloring technology can significantly improve production efficiency and image quality. This paper proposes an improved CycleGAN for anime line art coloring, constructing a generator structure that includes an encoder, transformer, and decoder, and designing a corresponding discriminator structure. Experimental results show that the improved model significantly outperforms the Pix 2 Pix model in terms of color restoration, detail preservation, and overall visual effect. Both subjective scores and objective evaluation metrics indicate that the improved model demonstrates higher accuracy and consistency in the image coloring task.
DOI:10.1109/CISAT62382.2024.10695217