Abstract: The difficulty in obtaining the “clear-blur” paired image enhancement dataset is the core problem restricting the development of image enhancement algorithms in underground coal mines, which indirectly limits the application and breakthrough of machine vision technology in the intelligent development of underground coal mines. Aiming at the problem of obtaining the current underground “clear-blur” paired image enhancement dataset, a construction method of underground image enhancement dataset based on generative is proposed: Based on the U–net deep neural network to learn the degradation features of dust and fog images caused by multiple coupling factors underground, the generation network is used to fit the high-quality degradation features of dust and fog to the clear underground image data, and the semantic consistency between the generated dust and fog images and the clear images is ensured through the cyclic consistency loss function. The “clear-blur” paired image enhancement dataset SDUcoal for the underground dust and fog image enhancement algorithm was constructed; A multi-dimensional index evaluation method based on “contrast, spectral energy distribution, NIQE, and Information entropy” was proposed to conduct a systematic quantitative analysis of the authenticity of the generated image dataset. A feasibility test method based on the “AODnet, Dehazenet” image enhancement algorithm was designed to verify and analyze the feasibility of the application of the generated image dataset. The experimental results show that in the evaluation based on multi-dimensional indicators, the dust and fog image data generated by this method and the real dust and fog image data underground: the contrast similarity reaches 82.04%, the similarity of high-frequency energy proportion reaches 82.06%, the NIQE similarity is as high as 92.16%, and the color histogram similarity reaches 65.81%. In the feasibility test: the SSIM reached 0.8363, 0.6210, 0.8406 and 0.8401 respectively, and the PSNR reached 20.62, 15.36, 25.29 and 24.51 respectively. Experiments prove that the “clear-blur” paired image enhancement dataset generated by this method can be used for the training of underground dust and fog image enhancement algorithms.