Unet, a deep learning architecture, has become one of the most widely used models for crack detection in the tunneling field. Although it performs well in overall crack image segmentation, it still has issues of limited feature expression capability and inaccurate segmentation. To address these problems, DTA-Unet was proposed based on dynamic convolution decomposition (DCD) and triple attention (TA). Firstly, the model used Unet as the baseline network and replaced traditional convolutions in the encoding-decoding process with DCD to enhance its feature extraction ability. Secondly, TA was combined with attention gate (AG) in the skip connections of the network, eliminating redundant information in spatial and channel dimensions to highlight the crack area. Finally, the proposed model was tested on crack datasets and compared with the conventional Unet model, image processing algorithms, and other deep neural network models in terms of detection performance on the datasets. The results show that it outperforms other advanced methods in crack detection performance. The proposed method is of significance to the maintenance of shield tunnel cracks.