Scene-guided YOLO framework with contextual feature fusion for solid waste detection in diverse real-world environments

Automated detection of solid waste in complex real-world environments represents one of the core challenges in environmental monitoring and smart city construction. Existing YOLO-based detection systems ignore scene-level semantic information when processing input images, resulting in insufficient cross-scene generalization and elevated false-positive rates in complex background regions. To address these issues, this paper proposes a novel Scene-Guided YOLO framework that integrates environmental contextual information into the YOLOv8 detection pipeline through a dual attention mechanism. Specifically, the framework incorporates a Feature-wise Linear Modulation (FiLM) module and a Scene-Enhanced Squeeze-and-Excitation (SE-Scene) module, implementing affine feature transformation and channel-wise attention recalibration respectively, enabling feature representations to dynamically respond to environmental context. Systematic experiments are conducted on a multi-environment dataset spanning five scene types and six recyclable waste categories, with three evaluation protocols—ablation study, convergence analysis, and noise robustness testing—used for comprehensive validation. Experimental results demonstrate that the proposed Scene-Guided YOLO achieves an mAP@0.5 of 66.7%, representing a 7.3 percentage point improvement over the baseline YOLOv8 model. Furthermore, the model exhibits significant robustness to scene label noise, consistently outperforming the baseline when the misclassification rate is below 35%. These results suggest that explicitly modeling scene–object relationships through contextual feature modulation can effectively improve waste detection accuracy and cross-environment generalization under the tested conditions.