Aluminum defect dataset.

Aiming at the current problems in the field of industrial defect segmentation, such as difficulty of obtaining a large number of defect samples, low recognition accuracy and lack of segmentation accuracy, a surface defect segmentation model for aluminum based on feature differentiation is proposed: FDTransUnet. First, the limited defective samples are effectively expanded by the feature differentiation data augmentation strategy to alleviate the overfitting problem caused by the insufficient sample. Second, the Transformer architecture is added by improving the U-net network, and the improved network combines the global self-attention mechanism of the Transformer and the hierarchical structure of the U-net, which can effectively extract the local and global information in the defect sample. Finally, a composite loss function is constructed to address the problem of unbalanced foreground and background sizes of defective samples and to improve segmentation accuracy. The experimental results show that FDTransUnet achieves 94.5% MPA and 89.7% Dice coefficient on the aluminum surface defect dataset. In the final generalization experiment, FDTransUnet is validated with other mainstream segmentation models on the steel surface defect dataset, and the experiment proves that the segmentation model has good generalization performance and robustness, and can be applied to different scenarios of industrial inspection.针对当前工业缺陷分割领域存在的大量缺陷样本获取困难、识别精度不足、分割精度缺失等问题，本文提出一种基于特征差异化的铝制表面缺陷分割模型FDTransUnet。首先，通过特征差异化数据增强策略对有限的缺陷样本进行有效扩增，以缓解样本不足引发的过拟合问题；其次，在改进U-net网络的基础上引入Transformer架构，将Transformer的全局自注意力机制与U-net的分层结构相结合，可有效提取缺陷样本中的局部与全局信息；最后，构建复合损失函数，以解决缺陷样本前景与背景尺寸失衡的问题，提升分割精度。实验结果表明，FDTransUnet在铝表面缺陷数据集上实现了94.5%的平均像素准确率（Mean Pixel Accuracy，MPA）与89.7%的Dice系数。在最终的泛化性能实验中，本文将FDTransUnet与其他主流分割模型在钢材表面缺陷数据集上进行验证，实验证明该分割模型具备优异的泛化性能与鲁棒性，可应用于多种工业检测场景。