The safe production of coal-fired power enterprises is an important foundation for ensuring the stability of energy supply and the safety of employees’ lives and property. Real-time and precise monitoring of the working scenes in mines is an effective means to prevent the occurrence of safety accidents. However, factors such as non-uniform lighting and high-concentration dust that are widespread in the underground environment have led to severe degradation of monitoring images, restricting the accuracy and timeliness of risk identification. To solve the above problems, this paper proposes a model-data dual-driven total variation model of depth maps (DGTV), embedding the spectral graph theory into the interpretable network structure to achieve depth restoration of low-quality images in coal mine shafts. The experimental results show that the DGTV algorithm proposed in this paper improves the peak signal-to-noise ratio (PSNR) by 6.00 dB and 1.93 dB respectively compared with the traditional classic denoising algorithm BM3D and the DnCNN based on deep convolutional neural network residual learning under medium light conditions. This method integrates spectral graph theory with deep networks to construct an end-to-end, trainable and interpretable image depth restoration scheme, providing high-definition visual data for emergency linkage, improving the efficiency of coal mine early warning and collaborative response, and ensuring personnel safety and facility stability.