Image content review method and system based on multi-layer scene graph structure

Abstract

This invention discloses an image content review method and system based on a multi-layer scene graph structure, belonging to the field of image content review technology. The method includes: segmenting an image to obtain entities and their attribute descriptions; clustering entities into several entity semantic clusters based on the entity's visual image features, the textual semantic features of the attribute descriptions, and the normalized spatial coordinates of the entity's center point, and analyzing the relationships between entity pairs within each cluster; identifying higher-order relationships between cluster pairs based on the overall semantic summary of each entity semantic cluster, and generating the overall narrative intent of the image; performing role recognition and referential analysis on the text in the image to obtain text-image relationships; integrating the generated data into a hierarchical graph structure, establishing thought chain prompts, and inputting it into a large language model for comprehensive reasoning. This invention enables a layer-by-layer, in-depth understanding of image content from local entities to global intent, accurately identifying complex and non-compliant semantics in images to be reviewed.

Images