The scene graph generation task focuses on optimizing relational representation and classification performance. To address the limitations of existing methods in high-order visual association mining and the fusion of visual and spatial information, this paper proposed a two-stage algorithm called HGM-VSE-NET (HyperGraph Modeling-Visual Spatial Enhancement Network) , the algorithm is based on visual-spatial cross-attention. This approach captures higher-order semantic associations between objects through hypergraph modeling, overcoming the limitations of traditional binary relations in expressing complex semantic dependencies. This paper constructs a visual self-attention mechanism to filter discriminative features, effectively suppressing redundant information while enhancing key visual context. Furthermore, this paper introduces a visual-spatial cross-attention module and spatial cosine similarity enhancement mechanism to achieve deep fusion of multimodal information, significantly improving the collaborative representation capability between spatial relationships and visual appearance. Experiments on the Visual Genome (VG) dataset demonstrate that the proposed method improves the mR@100 metrics by 3.2%, 4.5%, and 4.9% on the three subtasks of predicate classification, scene graph classification, and scene graph generation, respectively. This method significantly improves the recognition of low-frequency and medium-frequency predicates without compromising performance on high-frequency predicates. The results show that this method effectively alleviates the relationship classification bias caused by long-tail distribution and semantic ambiguity, and improves the discriminative ability and robustness of the model.