Abstract: Addressing the challenges of poor image quality and variable target scales in foreign object detection for under-ground mine belt conveyors, this study proposes a detection method that integrates an improved dark-bright channel prior dehazing algorithm with an enhanced YOLOv8 network. First, morphological closing operations are employed to refine the Otsu algorithm for accurate dark-bright region segmentation. Global atmospheric light is estimated through weighted fusion of dark and bright channel priors, while coarse transmittance maps are computed by applying corresponding channel priors to respective regions. Transmittance optimization is then achieved via Sigmoid function fusion and improved guided filtering. Second, YOLOv8 is enhanced through three modifications: designing an F-C2f module to optimize the backbone network, constructing C2f-E to strengthen feature extraction in the neck network, and implementing a dynamic detection head with Inner-WIoU loss function to improve detection head precision. Experimental results demonstrate that dehazing preprocessing improves baseline YOLOv8 detection accuracy by 2.2%. Compared with the original YOLOv8 model, the im-proved version achieves increases of 4.8%, 4.6%, and 3.8% in mAP@0.5, precision, and recall, respectively, while reducing floating-point operations and model parameters by 2.5% and 23%, respectively, thereby validat-ing the accuracy and efficiency of the proposed method for underground coal mine foreign object detection.