It is important to detect surface defects for controlling product quality and prolonging equipment life. However, detecting surface defects quickly and accurately is still a great challenge due to the complexity of the environment and surface defects. Aiming at the issue, this paper proposes a two-stage attention-based feature fusion network (TAFFNet) to make full use of each level feature for surface defect segmentation. Specifically, the network uses Resnet50 as the backbone network to obtain features, and then extracts multi-scale feature by the atrous convolution feature extraction module. In order to make the features at all levels contain more defect information, the attention-based adjacent feature fusion module is applied to fuse features with adjacent layers; then use the attention-based high-level feature fusion module to merge features with all upper layers, so all level features not only contain multi-scale context but also obtain more defect details. Finally, all features are cascaded together to achieve accurate segmentation of surface defects. In addition, TAFFNet uses a hybrid loss function to overcome blurry boundaries. The experimental results on three surface defects datasets (SD900, MT, and CFD) show that the proposed network outperforms other 13 methods in terms of PR curve, F₁, MAE, and mIoU.

检测表面缺陷对于控制产品质量和延长设备寿命非常重要。然而，由于环境和表面缺陷的复杂性，快速准确地检测表面缺陷仍然是一个巨大的挑战。针对该问题，本文提出了一种两阶段的基于注意力的特征融合网络（TAFFNet），以充分利用每一级特征进行表面缺陷分割。具体来说，该网络使用Resnet50作为骨干网络获取特征，然后通过atrous卷积特征提取模块提取多尺度特征。为了使各级特征包含更多的缺陷信息，应用了基于注意力的相邻特征融合模块，将特征与相邻层融合；然后使用基于注意力的高级特征融合模块将特征与所有上层进行融合，因此所有级别的特征不仅包含多尺度上下文，还可以获得更多的缺陷细节。最后，将所有特征级联在一起，以实现对表面缺陷的准确分割。此外，TAFFNet 使用混合损失函数来克服模糊边界。在三个表面缺陷数据集（SD900、MT 和 CFD）上的实验结果表明，所提出的网络在 PR 曲线、F₁、MAE和mIoU。