Adaptive Content Feature Enhancement GAN for Multimodal Selfie to Anime Translation

Yuanming Li, Youngsaeng Jin, Jeonggi Kwak, Dongsik Yoon, David Han, Hanseok Ko

Research output: Contribution to conferencePaperpeer-review

Abstract

With astonishing successes of GAN-based style transfer techniques, real-world photo translation to animation style images recently has attracted some interest. In particular, the transfer of a selfie to a cartoon has become quite popular as it can serve as a cartoon filter in social media. Unlike most of the Image-to-Image translation tasks, the selfie to anime task requires preserving the contour in the selfie image in the transfer process while it transforms other local characteristics into an animation style. Since the gap between the selfie domain and anime domain is quite large, as it can be imagined in the case of transforming a person into a cartoon animal like a mouse, developing an effective method remains a difficult challenge. In this paper, we propose an Adaptive Content Feature Enhancement Generative Adversarial Networks (ACFE-GAN) for a selfie to anime translation. By our model, the preservation of content features of selfie is improved. In addition to facial and hair contours, the shapes of worn items (e.g. hat and glasses) are also better preserved compared to existing methods. Our method also captures local features more accurately and selectively in translating them into the animation domain. Moreover, compared to the previous photo to anime translation models, we implement it with multimodal translation. Experiments on the selfie2anime dataset demonstrate that our method delivers superior performance in terms of selective preservation of content features.

Original languageEnglish
Publication statusPublished - 2021
Event32nd British Machine Vision Conference, BMVC 2021 - Virtual, Online
Duration: 2021 Nov 222021 Nov 25

Conference

Conference32nd British Machine Vision Conference, BMVC 2021
CityVirtual, Online
Period21/11/2221/11/25

Bibliographical note

Publisher Copyright:
© 2021. The copyright of this document resides with its authors.

ASJC Scopus subject areas

  • Artificial Intelligence
  • Computer Vision and Pattern Recognition

Fingerprint

Dive into the research topics of 'Adaptive Content Feature Enhancement GAN for Multimodal Selfie to Anime Translation'. Together they form a unique fingerprint.

Cite this