Abstract: To address the issue of insufficient illumination in coal mine shafts leading to low brightness and contrast in captured images, which in turn affects the feature extraction performance of coal gangue detection algorithms, this paper proposes an object detection method suitable for low-light environments underground. This method incorporates a low-light image enhancement module and an object detection module. First, the low-light image enhancement module improves the visual quality of raw, low-fidelity images and restores essential structural, textural, and chromatic features. Subsequently, the enhanced images are fed into the object detection module for accurate coal gangue recognition. Within the enhancement module, we introduce the MF-CG-LIME algorithm, an advancement over MF-LIME. This algorithm uses dark channel guidance and adaptive threshold segmentation to optimize brightness adjustment and weight calculation, thereby amplifying the contrast between coal gangue and background while preserving critical textures. To combat residual noise, we design the Enhanced Residual Attention Denoising Network (ERADNet), which leverages a reinforced residual architecture. By integrating cascaded residual blocks, skip connections, and channel-spatial attention mechanisms, ERADNet effectively targets noisy regions and salient feature channels, maintaining a precise balance between noise reduction and detail preservation. This substantially enhances key discriminative features for coal gangue detection. For object detection, we propose the ELW-YOLO module, specifically tailored for underground coal gangue recognition. Building upon the YOLOv10s framework, this module incorporates EfficientNetV1 components into the backbone to optimize the depth-width-resolution ratio, improving the extraction of subtle gangue features. The addition of the LSKA attention mechanism in the neck network further boosts adaptability and accuracy for multi-scale feature extraction. The WIoU loss function dynamically optimizes center point alignment and balances sample learning weights, improving the precision and stability of coal gangue boundary localization. Experimental results show that our image enhancement algorithm outperforms existing methods across all evaluation metrics. ELW-YOLO achieves an average precision of 90.8%, a 3.8% gain over original YOLOv10s, and delivers the best overall performance on images enhanced by our approach, reaching an average frame rate of 73.7 FPS. This enables real-time detection and provides technical support for target detection in the low-illumination, complex environments of underground coal mines.