Remote Sensing Image Scene Graph Generation Method Based on Knowledge Graph Enhancement and Relationship Filtering

Abstract. The generation of scene maps of remote sensing images is very important for the understanding of image depth, but its development is limited by the characteristics of wide image size, significant changes in target scale and dense distribution. In this paper, a scene graph generation method based on knowledge graph enhancement and relationship filtering is proposed. Firstly, the method takes the Reltr model as the backbone framework, constructs and integrates the domain knowledge graph, and uses the typical spatial relationships (such as orientation and topology) and semantic associations (such as functional constraints) of the encoded remote sensing targets as structured prior knowledge to guide the model to understand the more likely reasonable relationship patterns between targets. Secondly, the semantic information entropy mechanism was introduced to calculate the information entropy value of the probability distribution of the relationship class predicted by the model to quantify the uncertainty of the prediction. Based on this, an adaptive threshold is set to effectively filter and suppress low-confidence fuzzy or erroneous relationship predictions, focusing on mining deep complex relationships with high certainty between targets. Experiments on the STAR dataset show that the accuracy of the proposed method in the task of remote sensing scene map generation is significantly improved: the accuracy of the R@100, R@200, and R@500 reaches 23.1%, 25.6%, and 26.1%, respectively, and the mR@100, mR@200, and mR@500 reach 13.1%, 15.6%, and 17.1%, respectively, and the accuracy is better than that of the existing algorithms. This verifies the effectiveness of the method and provides a new and effective solution for the scene understanding of remote sensing images.