Abstract: Detection technologies for foreign objects on conveyor belts based on the YOLO series have produced fruitful research results. However, the neck network cannot enable direct feature exchange between distant feature layers, leading to problems such as missed detections and duplicate detections of small targets. Hyper-YOLO achieves high-order correlations in the neck network across layers and positions among feature layers, but it increases computational load and reduces sensitivity to high-frequency feature information, resulting in degraded feature extraction in noise-sensitive areas and shifted predicted bounding boxes. To address these issues, this study proposed a method for detecting foreign objects on coal mine conveyor belts based on an improved Hyper-YOLO. In the image preprocessing stage, the Dynamic Contrast Limited Adaptive Histogram Equalization (Dy-CLAHE) method was adopted, which introduced the Laplacian operator into the Contrast Limited Adaptive Histogram Equalization (CLAHE) framework to establish a dynamic mapping between noise level and contrast limited threshold. This effectively solved the problems of image detail loss and noise amplification in dusty environments. Hyper-YOLO was improved by using the Efficient Intersection over Union (EIoU) loss function to optimize the bounding box regression process, and enhanced localization accuracy. In the Mixed Aggregation Network (MANet), the Efficient Channel Attention (ECA) module was embedded into both deep and shallow layers to dynamically adjust channel weights through local cross-channel interaction, effectively balancing sensitivity to high- and low-frequency features and reducing the missed detection rate of small foreign objects. Meanwhile, the Simplified Spatial Pyramid Pooling-Fast (SimSPPF) module was used to reduce redundant computation, thereby improving inference speed without sacrificing accuracy. Experimental results showed that the improved Hyper-YOLO achieved an accuracy of 94.2% and a mAP@0.5 of 93.4%, representing improvements of 5.0% and 3.5%, respectively, compared to Hyper-YOLO. The number of parameters was 3.26×10⁶, recall was 87.7%, and detection speed was 158 FPS, which met the real-time detection requirements for underground coal mine foreign objects. In various detection scenarios, there were no missed or duplicate detections, and the predicted bounding boxes were better aligned with the foreign objects.