CKANST: CKAN-Based Arbitrary Style Transfer for Homogeneous Images

Résumé fourni par la source

Arbitrary style transfer has advanced in preserving content integrity and stylistic consistency. However, when the style image exhibits high homogeneity, popular methods often suffer from content leakage, which can lead to deviations in the content structure of the resulting stylized image. This paper proposes a CKAN-based arbitrary style transfer method for homogeneous style images, which effectively mitigates content leakage by leveraging CKAN’s powerful functional representation capability to extract and fuse features from both content and style images. In addition, the global feature fusion using Holistic Style Injector (HSI) is introduced to effectively model global information with CKAN operations and attention mechanisms. Experiments demonstrate that CKANST alleviates content leakage under homogeneous styles and delivers competitive performance compared with existing approaches, highlighting the potential of CKAN in image style transfer tasks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.