Abstract: In order to improve the detection speed for siltation of tunnel drainage holes by crystals and the accuracy of qualitative analysis of siltation degree, this paper explores and uses a convolutional neural network of semantic segmentation (DeepLab v3+ resnet18) to recognize the images of tunnel drainage holes. In this paper, the components in 230 drainage hole images were split into three categories: "crystal", "drainage hole wall" and "others". Furthermore, 138 images (60% of the total samples) were used to train the DeepLab v3+ resnet18 model, and the remaining 92 images (40% of the total samples) were used for image prediction. The results showed that the global accuracy based on this semantic segmentation model was up to 95%, and the prediction accuracy of crystals was over 75%, complying with the basic requirements for qualitative analysis of images of drainage hole siltation by crystals.In addition, this convolutional neural network of semantic segmentation was also self-programmed in MATLAB APP, so that staff could easily and conveniently detect (and predict) the images of drainage hole siltation by crystals.