Using SAM to Extract Semantic Boundaries in Complex Artistic Images to Assist in the Optimization of Graphic Layout Design

Although the Segment Anything Model (SAM) exhibits strong capability in visual boundary extraction, its limited high-level semantic understanding restricts its application to complex artistic images and intelligent design-oriented scenarios. This limitation is particularly evident in tasks requiring semantic interpretation beyond low-level edge detection. To address this issue, this study proposes a semantically enhanced segmentation framework that integrates textual prompts with design priors to improve image decomposition and structural understanding. The proposed framework provides end-to-end support from semantic image analysis to layout generation through interactive optimization and hierarchical boundary construction. Experimental results demonstrate that the method achieves an mIoU of 0.78 and a pixel accuracy of 0.95 on the ADE20K validation set. For visual center recognition and text avoidance, which are critical indicators of layout quality, the proposed approach attains a visual overlap of 0.82 and a text avoidance accuracy of 0.91, with an expert evaluation score of 4.2. Compared with baseline methods, the framework delivers superior segmentation performance, layout rationality, and operational efficiency. Beyond graphic design applications, the proposed semantic enhancement strategy provides an effective computational paradigm for multimodal visual information processing and electromagnetic imaging interpretation, offering valuable methodological insights for intelligent sensing systems and advanced image analysis in engineering applications.