X-ray tomography (CT) has become the main means of soil pore structure research, and CT image segmentation, as one of the key links, is directly related to the accuracy of pore analysis. Traditional image segmentation methods have limited accuracy, and in recent years deep learning techniques have been widely used in soil CT image segmentation. NAM (Normalization-based attention module) is an efficient attention mechanism that improves the design of spatial attention and channel attention sub-modules and can enhance the effect of deep learning model segmentation. In this study, four deep learning models, UNet, UNet++, DeepLabV3, and AttentionUNet with the addition of the NAM attention mechanism, were used to segment CT images of northeastern black soil and compared with traditional segmentation methods. The results showed that AttentionUNet achieved evaluation values of 88.08%, 0.96, 96.02%, 96.64% in the four indexes of intersection and concurrency ratio, F1 score, precision, and recall, respectively, and the precision was improved by 1.46 and 0.21 percentage points relative to UNet and UNet++, respectively, and compared with the traditional Watershed method, the precision and recall rate were increased by 101.13% and 33.54%. In conclusion, AttentionUNet model proposed in this paper can effectively improve the efficiency and precision of soil CT image segmentation, and provide a more efficient method for pore segmentation of soil CT images.