Abstract: To improve foreign object detection for underground coal mine belt conveyors under conditions of low image contrast, severe dust and haze interference, blurred target edges, and frequent missed detection of small-scale foreign objects, this study proposed a belt conveyor foreign object detection method integrating dark-bright dehazing and an improved YOLOv8 model. To address the problems of uneven dust and haze distribution and complex illumination in underground coal mines, an improved Otsu algorithm was employed to segment dark and bright image regions. The dark channel prior and bright channel prior were combined to estimate the transmission map and atmospheric light, respectively. A Sigmoid-weighted fusion strategy and an improved guided filtering method were introduced to optimize the transition between dark and bright regions and reduce edge artifacts and local distortion. The YOLOv8 model was improved by introducing the F-C2f module to reduce the computational complexity of the backbone network, the C2f-E module to enhance key feature extraction under complex backgrounds, and the DyHead dynamic detection head together with the Inner-WIoU loss function to improve the detection capability for multi-scale foreign objects and low-quality samples. A belt conveyor foreign object dataset was constructed for experimental evaluation. The results showed that the proposed dehazing algorithm achieved an average gradient of 31.251 3, representing an increase of approximately 347.6% compared with the original images. The average value of the Fog Aware Density Evaluator (FADE) decreased to 0.774 0, approximately 59.1% lower than that of the original images, indicating a significant improvement in image clarity. The improved YOLOv8 model achieved an mAP@0.5 of 94.2%, which was 5.6% higher than that of the original YOLOv8 model, with a detection speed of 52 frames/s, 7.9×10⁹ floating-point operations, and 2.7×10⁶ parameters. A foreign object detection experimental platform was established to conduct dynamic target detection experiments. The results showed that the proposed method integrating dark-bright dehazing and the improved YOLOv8 achieved detection accuracies of 90.7% and 82.0% for large foreign objects and rock bolts, respectively, representing improvements of 7.4% and 12.0% over the YOLOv8-based method.