Real-time analysis of tunnel muck images is considered significant for the intelligent level of tunnel boring machine (TBM) construction. Low contrast, large appearance differences, and severe adhesion between particles are often found in actually collected muck images. Great difficulties for accurate segmentation are caused by these issues. In order to solve these problems, an improved segmentation model named RAttU-Net based on deep learning was used to investigate the accurate segmentation of tunnel muck images. The U-Net encoder-decoder structure was utilized as the basis of this model. The recurrent recurrence of RU-Net was introduced to enhance the low-level feature reuse capability. The attention mechanism of AttU-Net was combined to strengthen the feature extraction of key areas. Furthermore, morphological erosion and seed filling algorithms were adopted in the post-processing stage. The effective separation of adhered muck particles was realized by these algorithms. A muck image dataset containing multiple types of TBM engineering scenarios was constructed to verify the model performance. The RAttU-Net was trained and tested. The results show that a maximum pixel accuracy (PA) of 0.921 and a maximum F1-score of 0.884 are achieved by the proposed method in the region segmentation task, compared with U-Net, AttU-Net, and RU-Net baseline models. A 60% Hausdorff distance (60-HD) of 2.98 mm is obtained for muck particle segmentation. A 95% Hausdorff distance (95-HD) of 3.54 mm is observed. In addition, a processing time of only about 2.85 s is required by the model for a single-frame full-size image. The predicted gradation curve is highly matched with the actual distribution. It is concluded that reliable technical means are provided by the research results for the real-time identification and analysis of muck morphology during TBM construction. The development of intelligent tunnel boring construction is actively promoted by this technology.