A Robust RT-DETR-Based Method for Complex Self-Service Buffet Scene Detection

Object detection in buffet-style environments is highly challenging due to densely stacked tableware, frequent occlusions, strong illumination reflections, and substantial visual similarity across categories, all of which undermine the robustness of existing detectors. To address these issues, this paper proposes an improved real-time detection transformer–based model with a lightweight design while significantly enhancing multi-scale feature representation. First, a re-parameterized stem module is introduced to strengthen shallow texture extraction with negligible computational overhead. Second, a dynamic multi-kernel refinement module is developed to enrich directional texture modeling and cross-scale semantic aggregation. Furthermore, a heterogeneous-kernel feature pyramid network is constructed by integrating adaptive multi-scale fusion, multi-kernel fusion nodes, and a lightweight upsampling strategy to improve cross-level feature consistency and mitigate aliasing caused by conventional upsampling. Experimental results on a self-constructed buffet-scene dataset demonstrate that the proposed method improves mAP50 and mAP50:95 by 2.6% and 1.9%, respectively, while reducing parameters and GFLOPs by 42.6% and 42.3%, and increasing inference speed to 103.1 FPS. On Dota v1.0 and SkyFusion data sets, the small target detection ability has also been improved. The substantial reductions in computation and model size further confirm the effectiveness and practical value of the proposed approach for complex catering scenarios.