Existing urban object detection and analysis methods still lack effective mechanisms to compute and rank semantic relevance between objects in urban scenes. This deficiency limits the recognition accuracy in practical applications and affects the efficiency and precision of subsequent processing. This paper proposes a vision model-based approach for image object relevance analysis, aiming to evaluate inter-object correlations in images by integrating scene knowledge graphs with target relevance analysis.First, we systematically conduct ontological modeling of urban scenes to construct a cityscape knowledge graph. Building upon this framework, we introduce an algorithm combining the visual Relevance Assessment Model (Recognize Anything Model:RAM) with personalized PageRank to calculate semantic relevance between urban scenes and their constituent objects. Based on the analytical results, we implement preference ranking for targets, prioritizing key objects with higher relevance weights to enhance system efficiency and accuracy.Experimental results demonstrate that the proposed method outperforms conventional object detection approaches in recognition accuracy, task relevance matching degree, and computational efficiency, validating its effectiveness and superiority in complex urban scenarios.