Visual relationship detection is a fundamental task in computer vision, focusing on identifying spatial and semantic relationships between objects within an image. While existing approaches primarily target human-object interactions, limited attention has been given to pairwise relationships among densely distributed objects—a setting common in real-world scenarios. Moreover, few models offer both high accuracy and efficient inference. To address these limitations, we propose a DETR-based visual relationship detection model that incorporates query selection to effectively model object-object interactions. Our approach simplifies the detection pipeline and enables end-to-end prediction without relying on region proposals or post-processing. In addition, we explore the application of this model in the field of document image understanding by constructing and annotating a dataset of office diagrams, including flowcharts, to support structured visual-text parsing. A series of experiments are conducted to evaluate the model’s performance from multiple perspectives. Results demonstrate that the proposed method achieves accurate and robust relationship detection, even in images with high object density, while maintaining computational efficiency. These findings highlight the model’s potential for broader applications in intelligent document analysis and digital office automation.