A Coal Gangue Debris Detection Method Based on YOLOv8-FAPF

In view of the issues faced by the YOLOv8 model in the coal gangue debris detection task, such as insufficient feature representation ability due to dense background interference, target shape heterogeneity, missing multi-modal information (only grayscale input), and dynamic environmental noise (e.g., lighting, fog), which result in false positives, missed detections, and low detection accuracy, an improved YOLOv8-FAPF model is proposed. The feature map P2, which is obtained by downsampling the original image by a factor of 4, is introduced as the input to the neck network. Its larger receptive field and richer low-level features enhance the detection performance of small objects. By optimizing Path Aggregation Network into an Adaptive Pixel-wise Weighted Fusion (APWF) approach, it performs adaptive weight assignment and dynamic fusion of multi-scale features at the pixel level, while adding additional channels for information fusion. This significantly enhances the robustness of object detection in complex scenarios. The traditional interpolation sampling method is improved by adopting a frequency-aware feature fusion approach. By introducing the FreqFusion module, the upsampling process is optimized to enhance feature consistency and clarify object boundaries, further improving the performance of object detection, particularly in the task of detecting coal gangue debris with complex backgrounds and multi-scale targets. Ablation and comparative experiments demonstrate the effectiveness and superiority of the proposed YOLOv8-FAPF model. Experimental results show that the improved YOLOv8-FAPF achieves a mAP50-95 of 93.3% on the selfconstructed coal gangue dataset, a 3.7 percentage point improvement over the baseline model, while maintaining a real-time processing speed of 2 ms/frame. Comparative analysis with mainstream detection models shows that YOLOv8-FAPF outperforms Faster R-CNN (+10.3%), DETR (+14.2%), and Deformable-DETR (+8.2%) in terms of mAP50-95. The YOLOv8-FAPF exhibits the strongest adaptability and the best overall detection performance in complex environments containing contaminated coal gangue, providing an effective technical solution for real-time debris detection in complex industrial settings. Comparative analysis with mainstream detection models shows that YOLOv8-FAPF surpasses Faster R-CNN (+10.3%), DETR (+14.2%), and Deformable-DETR (+8.2%) in the mAP50-95 metric. YOLOv8-FAPF demonstrates the strongest adaptability and the best overall detection performance in complex environments, providing an effective technical solution for real-time coal gangue debris detection in complex industrial settings.