Industrial surface defect segmentation is crucial for addressing quality control issues in production and manufacturing. However, limitations such as insufficient sample sizes, low data utilization, and a lack of generalization of the models persist. To address these issues, this paper introduces a universal surface defect segmentation model, fusion attention \(\text {U}^{2}\) -net(FA \(\text {U}^{2}\) -net), which utilizes an attention mechanism to focus intensely on features, including spatial structures. By incorporating this attention mechanism and multiple loss functions, the model enhances its focus on critical features and demonstrates more robust generalization capabilities. This model outperformed other comparison models in various tests. Specifically, on datasets with uneven sample distributions, our model achieved a mean intersection over union (MIoU) of more than 80%, and on other datasets, it reached an MIoU of more than 85%. These results demonstrate the robustness and broad applicability of this model. Finally, through ablation studies, this paper validated the rationality and effectiveness of each component within our network.
