Abstract: The image semantic segmentation technology is introduced to segment the object in the sub-illumination environment of mine, and the image is divided into two categories: original clear image and sub-illumination image. The image enhancement method based on deep learning is used to replace the enhanced details of the images taken under sub-illumination conditions, and then the data set is expanded by monogram transformation, and then the standard data set of semantic segmentation of mine roadway images is constructed. A lightweight encoding-decoding structure network based on the self-attention mechanism was proposed, which was based on DeepLab V3+ coding-decoding network. In the encoding structure, the deep and shallow semantic feature information of the mine image was extracted, and the deep semantic feature information was activated by the lightweight self-attention mechanism module, and shallow semantic feature information was directly sent to the decoder. The deep semantic feature information and shallow semantic feature information were spliced in the decoding structure, the original image size was restored, and the segmentation result was output. Compared with the traditional algorithm, the experimental results show that: in terms of network complexity, for the 3-channel 512×512 image, the network theoretical computation cost of the algorithm is only 48.80 G FLOPs and the parameter number is only 11.90 M; in terms of network segmentation accuracy, the average intersection ratio is 76.50% and the average pixel accuracy is 87.75%, leading other mainstream networks; in terms of speed, the speed of an image can reach 0.032 s, meeting the requirements of lightweight networks.