Objective The roughness of a concrete joint surface significantly affects its shear performance and bonding strength. Therefore this study designs a deep learning model capable of identifying the roughness grade of concrete joint surfaces. The model enables a more intuitive and rapid assessment of joint surface roughness. Methods First ResNet50 was used as the backbone network to identify the roughness grades of the collected depth maps of concrete joint surfaces. Thus a deep learning model for roughness identification based on this backbone network was obtained. Then to further improve the identification accuracy of the model three different embedding strategies were adopted. The convolutional block attention module CBAM its channel attention component and its spatial attention component were respectively embedded behind the first convolutional layer of ResNet50. The parameters of the backbone network were kept unchanged during this process. As a result three deep learning models that combined convolutional and attention mechanisms were obtained. Results Simulation results showed that the identification accuracy of ResNet50 was 86. 99% compared with 82. 88% for GoogLeNet and 83. 56% for AlexNet. When three different attention mechanisms were embedded into ResNet50 the model achieved accuracies of 60. 27% after embedding the full CBAM 93. 84% after embedding the channel attention and 91. 78% after embedding the spatial attention. Conclusion Compared with GoogLeNet and AlexNet ResNet50 demonstrates higher identification accuracy. Among the attention-enhanced models the ResNet50 embedded with the CBAM channel attention mechanism achieves the highest accuracy. Therefore this model is more effective in identifying the roughness grade of concrete joint surfaces.