Underground coal mining faces hazards like gas explosions, coal dust explosions, and fires, underscoring the need for safety robotics based on 3D reconstruction. While essential for precise navigation and detection in complex environments, traditional methods fall short in data quality, accuracy, and cost. To address this, a 3D reconstruction method called 2DGS-DbTrans for underground coal mine tunnels was proposed, which is based on pure vision-based 3D reconstruction technology. To improve image resolution, a Transformer module was designed to enhance the input images, consisting of two core components: the multi-head depth convolutional axis attention mechanism and the deep convolutional gated network. In the processing workflow, sparse point clouds were first generated using Colmap, and the underground mine tunnel environment was represented by 2D Gaussian surfaces, where each Gaussian surface contains the coordinates, color, size, and orientation information of the objects. In addition, two loss functions were defined: the color loss function and the road smoothness loss function. Experimental results show that the 2DGS-DbTrans method outperforms other methods in 3D reconstruction of underground coal mine tunnels.