Abstract: [Purpose/Significance] Aiming at the semantic missing and incomplete problems in the process of image organization and retrieval, a framework for semantic description of images in social media is proposed to enrich the existing theoretical system of image description, improve the efficiency and utilization of image retrieval, and provide a reference for the realization of the automatic semantic annotation of images. [Method/Process] First, we conducted a survey and analysis of research progress related to image description both domestically and internationally, summarizing the existing theories of image description and annotation, metadata specifications, and related technical methods. Second, based on the image metadata standards and the theory of hierarchical and categorical description of image features, we constructed a semantic description framework for social media images, focusing on seven layers: external feature layer, content layer, object layer, relationship layer, scene layer, event layer, and emotional layer. We also elaborated in detail the various semantic layers and their interrelationships. Finally, we verified the feasibility of the image semantic description framework by describing the examples of character images and landscape images. [Results/Conclusions] The results of the descriptive examples of character images and landscape images indicate that the image semantic description framework can eliminate the "semantic gap" in image description through semantic associations between different layers, and achieve a multi-faceted, multi-dimensional, and multi-level structured and semantic description of the external and content features of images. It has strong portability and flexibility. However, there are also certain limitations and areas for improvement in this paper: 1) Based on the image semantic description framework proposed in this paper, a prototype system based on image annotation needs to be developed; 2) The images posted by users on social media are closely related to the situation, and they are more likely to express emotions. In the future, more research on the semantic layer of images can be conducted based on the text information posted by users; 3) Future research can further explore the application of deep learning in image and text fusion to achieve more accurate event and emotion recognition. By constructing a more complex neural network structure, the event and emotion information in the image can be deeply mined and fused; 4) When describing images, the study should pay attention not only to static visual features, but also to consider the dynamic course of events. Future frameworks could attempt to combine static and dynamic information to provide richer, more vivid descriptions of images.

Key words: semantic description framework, image feature, semantic annotation, Sora