Surface defect inspection is a crucial step to ensure the quality of magnetic tiles. Recently, deep learning methods have shown excellent performance on many vision tasks. Some deep learning-based methods have been applied to the surface defect inspection of magnetic tiles as well. However, related methods are based on supervised learning, which requires plenty of labeled samples to train deep neural networks. In industrial application scenarios, the annotation of large labeled datasets is extremely expensive, time-consuming, and error-prone. A semi-supervised learning method based on pseudo-labeling is proposed in this paper to address the problem of surface defect classification of magnetic tiles with limited labeled samples. The proposed method consists of two models: the teacher model and the student model. The training procedure is divided into two stages: pseudo-label generation and student model training. In the pseudo-label generation stage, the teacher model parameters and the pseudo-labels of unlabeled samples are alternatively optimized based on the idea of transductive learning. Curriculum learning is employed to reduce the impact of label noise so that high-quality pseudo-labels can be obtained. In the student model training stage, labeled samples and unlabeled samples with pseudo-labels are jointly used to train the classifier, with mixup to achieve information fusion and regularization. The experimental results show that the proposed method outperforms the supervised-only and semi-supervised baselines. With only 4.4% of labeled samples in the training set, the proposed method can still achieve the defect classification accuracy of 90.13%.

表面缺陷检测是确保磁砖质量的关键步骤。最近，深度学习方法在许多视觉任务上都表现出了出色的表现。一些基于深度学习的方法也已应用于磁砖的表面缺陷检测。然而，相关方法是基于监督学习的，这需要大量的标记样本来训练深度神经网络。在工业应用场景中，大型标注数据集的标注极其昂贵、耗时且容易出错。针对标记样本有限的磁砖表面缺陷分类问题，提出了一种基于伪标记的半监督学习方法。所提出的方法由两个模型组成：教师模型和学生模型。训练过程分为两个阶段：伪标签生成和学生模型训练。在伪标签生成阶段，教师模型参数和未标记样本的伪标签基于转导学习的思想交替优化。课程学习用于减少标签噪声的影响，从而可以获得高质量的伪标签。在学生模型训练阶段，联合使用带伪标签的标注样本和未标注样本训练分类器，通过mixup实现信息融合和正则化。实验结果表明，所提出的方法优于仅监督和半监督基线。在训练集中只有 4.4% 的标记样本的情况下，所提出的方法仍然可以达到 90.13% 的缺陷分类准确率。