Abstract: To address the problem of large differences in foreign object sizes, complex backgrounds, and poor detection performance of small and slender targets in coal mine conveyor belts, a foreign object detection method based on an improved YOLOv11n model was proposed. The core improvements of the YOLOv11n model included three aspects: first, a Scale Sequence Feature Fusion (SSFF) module was introduced into the neck network to enhance the effective capture and fusion of information at different scales through sequential scale interaction; second, a parallel C3K2PPA module was constructed, introducing a Parallelized Patch-Aware Attention (PPA) module in the spatial dimension to highlight key region representations and improve recall; third, a Band-Aware Contrast Fusion (BACF) module was designed at the lateral fusion layer of the same scale. By combining belt-directional priors and high-pass edge indicators, and replacing simple concatenation with pixel-wise gating, the module suppressed periodic background noise along the belt direction and enhanced cross-branch differences without increasing the number of channels, thereby improving the model’s discriminative capability and robustness under complex working conditions. The experimental results showed that the precision and recall of the improved YOLOv11n model reached 0.914 and 0.892, respectively, with mAP@0.5 and mAP@0.5:0.95 values of 93.1% and 62.2%, showing significant improvement over the original YOLOv11n and outperforming mainstream lightweight models such as YOLOv5s, YOLOv8n, and YOLOv10n in accuracy and robustness. The inference speed of the model reached 96 frames per second, indicating high real-time performance and efficient execution in coal mine conveyor belt foreign object detection tasks. Heatmap analysis showed that the improved YOLOv11n model effectively enhanced the target-area focusing capability, reduced redundant bounding boxes, and improved the detection accuracy of small targets.