Mine video surveillance technology has significant advantages in promoting coal mine safety, high efficiency, and unmanned mining. However, the poor working conditions of cameras in coal mines have led to serious degradation of image quality. For this reason, multi-source fusion processing of mine monitoring video will solve the current problems and be helpful to promote the intelligent development of coal mines. In view of the low computational efficiency of existing image fusion algorithms and poor timeliness, the fusion images acquired by the existing image fusion algorithms have problems such as false targets, fuzzy targets, halo occluded targets, etc., which cannot meet the needs of mine video surveillance. This paper proposes a mine dual-band image fusion algorithm using multi-scale and adaptive Gaussian difference transform combined with rectified non-linear unit and VGG-16. A source image decomposition model based on multi-scale and adaptive difference of Gaussian is designed. This model decomposes infrared and visible images into basic images and detailed images. Among them, the basic image represents the approximate components of the source image, reflecting the general features of the field of view. The detailed image represents the detailed components of the source image, including detailed information such as edges and textures, and is also the most sensitive part of human eye recognition and machine vision. To eliminate the interference part of the light source in the visible basic image and improve the overall contrast and information richness of the fusion image in the underground mine, a rectified non-linear unit function is constructed. The rectified non-linear unit function makes the weight of the infrared basic image automatically adjust with the gray level of the visible basic image, and the “weighted average” basic image fusion strategy is adopted to obtain the fused basic image that eliminates the interference of the light source and retains the general features of visible and infrared images. Meanwhile, the pre-trained VGG-16 network model is used to extract the 4-layer depth features of the detailed image, and the l₁-norm and Gaussian operator are used to sequentially obtain the saliency maps corresponding to the 4-layer depth feature. After obtaining 4 pairs of fused images with different depth features through pooling inverse operation and weighted fusion, the fused detail image is obtained by the “maximum value selection” method. The fusion basic image and detail image are reconstructed to obtain the final fusion image. To verify the effectiveness of the proposed algorithm, the experiment selected the source images of the coal mine in four different scenes, combined five typical image fusion algorithms for subjective analysis, and used five quality evaluation indicators of fusion image and average running time for objective evaluation. The experimental results show that the proposed algorithm can eliminate the interference of artificial light sources and obtain fused images of underground mines with clear scenes and salient features, and the fused image is more in line with human visual characteristics. At the same time, it improves the fusion quality and fusion efficiency of heterogeneous images, which is conducive to the further analysis and processing of images. Compared with the other five typical algorithms, the proposed algorithm is more robust. It not only overcomes the shortcomings of traditional algorithms that cannot extract image depth features but also makes it easier to completely eliminate light source interference and obtain more comprehensive, reliable, and rich scene information. In addition, the proposed algorithm can be used for the intelligent analysis of multi-source images of mines and remote monitoring on the ground. It can also be used to eliminate the problem of artificial light source interference in underground space, underground engineering or night road video surveillance images.