CrePoster: Leveraging multi-level features for cultural relic poster generation via attention-based framework

Abstract

Integrating multi-level features of cultural relics into aesthetic posters suitable for social sharing can enhance cultural value dissemination. However, this task meets two major challenges: (i) generating professional captions for cultural relics that encompass color, shape, form, and metaphor details; (ii) combining designers’ expertise and cultural relics’ unique aesthetic features to create visually appealing posters through layout, color, and font selection. Existing methods for poster generation primarily target merchandise or scientific publications. Constrained by product style or traditional design rules, which are unsuitable for cultural relics with multi-level aesthetic features. In this work, we propose CrePoster, an attention-based Cultural relic Poster generation framework that incorporates multi-level feature extraction. Taking Chinese cultural relics as the case study, after the photos are uploaded, CrePoster leverages a large-scale pre-trained image segmentation network to obtain the critical object. Subsequently, a multi-level feature extraction-based caption generator is utilized to generate professional captions. Afterward, an attention-based dual-scale fusion network is employed to represent the aesthetic characters and guide the layout matching. Compared with existing methods, CrePoster can generate higher-quality captions and posters with more aesthetic value.