Abstract: The semantic description and annotation of scientific document images are essential for enhancing knowledge mining and intelligent information management of academic literature. Based on metadata structures and information requirements of academic images, this paper proposes a multi-level semantic description framework for scientific paper images (SDF-SLI), aimed at achieving precise semantic parsing of scientific document image content in the context of AI4Science. The framework analyzes scientific document images through four interconnected layers—basic identification layer, content layer, semantic layer, and relationship layer—which examine basic information, visual content, semantic connotation, and hierarchical relationships. This research also establishes a comprehensive ontology model that systematically maps semantic relationships between these framework layers, utilizing multimodal large language models for implementation. Through empirical case studies in library and information science literature, the framework demonstrates its practicality and effectiveness. The paper further discusses key challenges and future opportunities in semantic description of scientific paper images, offering insights into the intelligent development of scientific knowledge management.