Research on Automatic Segmentation and Layout of Poster Theme Visual Elements Using Text-Prompt-Driven 3D Base Model (SAT) and Contrastive Learning

Abstract

Automatic segmentation and layout of poster visual elements require joint modeling of visual structure, theme semantics, and spatial configuration. To overcome weak theme understanding, unclear element boundaries, and limited semantic relevance in existing poster-design automation methods, this study proposes a text-prompt-driven segmentation and layout framework based on a SAT spatial-semantic foundation model and contrastive learning. A text-image-layout triplet dataset is constructed, containing poster theme descriptions, pixel-level masks, bounding boxes, and element category labels. The SAT encoder extracts multi-scale visual features, while a BERT-based prompt encoder projects theme semantics into a high-dimensional representation to guide visual attention toward theme-relevant regions. A contrastive-learning module constructs semantic association matrices among elements, and an attention fusion module outputs element masks and layout coordinates. Experiments on a self-constructed poster dataset show that the proposed method achieves an average IoU of 0.870, a layout rationality score of 8.6/10, and an inference time of 0.30 s per image. Ablation experiments confirm the contribution of text prompting, semantic guidance, and contrastive learning. The framework supports visual signal processing, multimodal feature alignment, and spatial layout optimization in intelligent design systems.