AutoCaCoNet: Automatic Cartoon Colorization Network Using Self-Attention GAN, Segmentation, and Color Correction

Seungpeel Lee, Eunil Park; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2024, pp. 403-411

Colorization is a captivating research area within the realm of computer vision. Conventional methods often rely on object-based strategies, necessitating access to extensive image datasets. However, recent advancements in deep neural networks have illuminated the feasibility and practicality of automating image colorization tasks. This study introduces a pioneering automatic cartoon colorization network named Automatic Cartoon Colorization Network using self-attention GAN, segmentation, and color correction (AutoCaCoNet), harnessing the power of a conditional generative adversarial network (GAN) coupled with self-attention, segmentation, and color correction techniques. The ensuing experimental results, meticulously presented through both qualitative and quantitative assessments, underscore the significance of AutoCaCoNet. This significance is particularly evident when applied to a real-world cartoon dataset, surpassing the performance metrics of preceding research endeavors. Furthermore, the findings from a user survey, encompassing both ordinary users and expert groups, consistently award AutoCaCoNet the highest scores. We are pleased to announce the availability of our codebase and dataset to the public, encouraging further exploration and advancement in this domain.