시각적 주목도 맵 기반 포스터 디자인 의도 영역 예측

Abstract

Accurately detecting user design intent in poster design is a crucial element for automated layout generation systems. Existing methods have shown limited performance in complex poster compositions by relying on single input images. This study proposes a Cross-Attention-based dual-path UNet model that effectively fuses original poster images with visual attention hints. The proposed model processes each input through two independent encoders and applies Cross-Attention mechanisms at multiple scales to model interactions between the two inputs. Experimental results on the PKU dataset achieved superior performance with IoU 0.88 and Dice coefficient 0.93, demonstrating significant improvement over existing single-input-based methods. This research presents a novel approach in the field of design intent detection and is expected to serve as a core component for future automated poster generation systems.