PGFD-YOLO: A dual-modal object detection framework with progressive gated fusion and foreground-guided distillation

Effective waste classification is essential for improving resource recovery and reducing landfill burden. However, existing methods predominantly rely on single RGB imaging, which is insufficient for distinguishing overlapping, transparent, or similarly colored waste items. This study proposes a dual-modal waste detection model, PGFD-YOLO, based on YOLOv11, which integrates RGB and depth information without requiring dedicated depth sensors. Depth maps were generated from standard RGB images using monocular depth estimation (Depth Anything V2). Three architectural innovations were introduced: LKC3-F for lightweight dual-stream feature extraction via partial convolution, QAT-PSA for high-level semantic refinement via quad-enhanced attention, and PCGF for stable cross-modal integration via progressive zero-initialized gating. A foreground-guided cross-scale knowledge distillation strategy further improved the accuracy without increasing the inference cost. Experiments on the TACO-10 dataset with 10-fold cross-validation show that PGFD-YOLO achieves mAP₅₀ of 22.3±0.8% and mAP_50:95 of 16.5±0.7%, improving over the single-modal baseline by 7.7% and 7.2% absolute, and outperforming naive early and late fusion strategies by 8.1% and 6.4%, respectively. A projected deployment analysis estimates approximately 70%-75% labor cost reduction and $91,700-167,400 annual recyclable recovery value from a medium-scale facility. The model operates at 73.6 FPS, thereby meeting the real-time waste sorting requirements.