In response to the issues of low coal gangue recognition accuracy, missed detections, and false detections due to the complex working conditions of low illumination, high blur, and occlusion in underground coal mines, an ELS-YOLO coal gangue detection model is proposed. The ELS-YOLO model is based on the YOLOv10s model. Firstly, the EfficientNetV1 module is introduced into the backbone network, which expands the dimensions proportionally, balancing the scaling of the network's depth, width, and resolution, thus improving the performance of the convolutional neural network. Secondly, the LSGE attention module is introduced in the neck network, which significantly enhances the image feature extraction quality and efficiency through strategies such as local feature enhancement, multi-scale information fusion, and spatial feature aggregation. Finally, the SPPELAN module, which has an efficient local aggregation network, is selected to improve the model's detection performance when handling targets of different sizes. Experimental results show that the average detection accuracy of the ELS-YOLO model reaches 89.6%, which is 3.0% higher than the YOLOv10s model; the average detection speed is 81.30 FPS, fully meeting the real-time detection requirements of coal gangue in underground coal mines. Compared to YOLOv5s, YOLOv7-Tiny, YOLOv8s, and YOLOv9s, the ELS-YOLO model demonstrates the strongest adaptability to complex coal mine environments and the best overall detection performance, providing technical support for the intelligent and efficient sorting of coal gangue.