Abstract: Large lump material mixed on conveyor belts during belt conveyor operation can easily cause transfer point blockage and conveyor belt damage, posing a potential threat to safe system operation. To address the missed and false detections that may occur in vision-based large lump material detection models under complex underground conditions, a robust detection method for large lump material on belt conveyors under complex visual degradation conditions was proposed. Based on the imaging degradation mechanism, a detection-driven adaptive image enhancement module was constructed. A learnable convolutional unit was embedded at the front end of the object detection network to replace the explicit modeling process in traditional Retinex that depended on manually set parameters, thereby realizing adaptive enhancement of degraded images. This mechanism completed end-to-end joint optimization under the constraint of the detection loss function and did not require manual setting of enhancement parameters or distinction between degradation types. YOLOv11 was selected as the baseline detection framework, the adaptive image enhancement module was introduced, and the SIoU loss function was used to assist detection performance optimization through a bounding box regression strategy with stronger geometric constraints. Experimental results showed that the improved YOLOv11 model had good detection stability and robustness under foggy, low-illumination, and mixed conditions. Compared with the original YOLOv11 model, detection accuracy was improved by 1.3%, and missed detections in degraded images were significantly reduced. In terms of model complexity and real-time performance, the adaptive image enhancement module introduced only 0.1×10⁶ additional parameters and 6.3 GFLOPs of computational overhead. The inference speed of the improved YOLOv11 model reached 106 frames/s for images and 42 frames/s for videos, indicating that the proposed method maintained good real-time performance while improving detection accuracy.