Abstract: Belt conveyors are key equipment in mine continuous transportation systems. During operation, large bulk materials mixed within the conveyed material can easily cause transfer point blockage and belt damage, posing potential threats to system safety. To address image degradation caused by fog and low illumination in underground environments—which often leads to missed and false detections in vision-based large object detection models—this paper proposes an online detection method for large bulk materials on belt conveyors under complex working conditions. Based on the imaging degradation mechanism, a detection-driven adaptive image enhancement mechanism is constructed by embedding learnable convolutional units into the front-end of the object detection network, replacing the explicit modeling process in traditional Retinex methods that relies on manually set parameters, thereby achieving adaptive enhancement of degraded images. The proposed mechanism is jointly optimized in an end-to-end manner under the constraint of the detection loss function, eliminating the need for manual parameter tuning or degradation type classification. For the detection framework, YOLOv11 with superior performance is adopted as the baseline model, into which the aforementioned lightweight learnable feature enhancement module is integrated. In addition, the SIoU loss function is employed to introduce stronger geometric constraints in bounding box regression, further improving detection performance. Experimental results demonstrate that the proposed method exhibits strong stability and robustness under foggy, low-illumination, and mixed degradation conditions. Compared with the original YOLOv11 model, the proposed approach improves detection accuracy by approximately 1.3% while maintaining a real-time inference speed of 106 frames per second under the same hardware conditions. Furthermore, compared with YOLOv5, YOLOv8, and YOLOv10, YOLOv11 achieves superior overall performance in large bulk material detection tasks. Compared with fixed image enhancement methods such as histogram equalization and Retinex, the proposed adaptive enhancement mechanism demonstrates stronger environmental adaptability and greater engineering application value under complex working conditions.