Abstract: Due to the complexity of the spatial environment and poor lighting conditions in underground coal mines, the images obtained by vision devices are prone to problems such as insufficient contrast and poor texture details, which seriously affect the reliability of the work of vision devices and limit further image-based intelligent applications. To improve the contrast of low-illumination images in underground mines while enhancing their texture details, a deep neural network-based low-illumination image enhancement model is proposed, which contains three sub-networks, namely, decomposition network, illumination adjustment network and reflection reconstruction network. The decomposition network decomposes the underground coal mine image into light and reflection components; the light adjustment network effectively reduces the parameters of the model using depth-separable convolutional structure and strengthens the feature extraction ability of the network; in addition, the MobileNet network structure is introduced to further lighten the light adjustment network while maintaining its feature extraction accuracy and effectively realizing the contrast adjustment of light components; the reflection reconstruction network introduces a residual network structure to improve the contrast adjustment of light components. Finally, the processed illumination and reflection components are fused based on Retinex theory to obtain enhanced images, which achieve contrast enhancement and detail enhancement of underground mine images, overcoming the problems of detail loss, blurred edges, and lack of contrast and clarity of the enhanced image that exist in existing enhancement algorithms. Numerical experiments show that the proposed model can effectively enhance the texture details of the image while improving the contrast of underground mine images, and has good stability and robustness, which can well meet the needs of low-light image enhancement in coal mines.