To meet the continuously increasing quality standards in industrial man-ufacturing processes, machine vision-based surface inspection systems arewidely used for checking surface quality and removing defective products.Automated Surface Inspection (ASI) algorithms are the core of the soft-ware for these systems, and their accuracy and effectiveness are crucial forthe success of these visual inspection systems. Compared with traditionalimage processing techniques-based ASI algorithms, deep learning-basedASI algorithms enable highly accurate automatic defect detection andare widely adopted by researchers in academia and industry. With alarge amount of labeled data, supervised deep learning can extract fea-tures from training samples and detect different types of surface defects.However, obtaining a large number of labeled samples is costly and oftenimpractical. Therefore, this paper proposes a semi-supervised learningapproach based on FixMatch, which requires only a limited numberof labeled samples but achieves performance comparable to supervisedlearning. This paper further introduces an uncertainty-adaptive marginmechanism to balance the learning weights between labeled and unla-beled samples, enhancing the model’s generalization ability under limitedlabeled data and preventing overfitting. Experiments on two public sur-face defect datasets (DAGM and NEU) and one industrial defect dataset(CCL) were conducted. The experimental results show that the proposedmethod, even with a small number of labeled samples, achieves high accu-racy rates of 99.83%, 99.81%, and 97.99% on the DAGM, NEU, and CCLdatasets, respectively, outperforming some supervised learning methodsand benchmark semi-supervised learning methods.