Semi-Supervised Surface Defect Segmentation via Contrastive Learning

Quality inspection of industrial products is a critical task to ensure that products meet the required standards. Although deep learning-based detection models have achieved significant progress in recent years, the limited availability of samples in real industrial environments still constrains the generalizability of these models. In this paper, we propose a novel semi-supervised defect segmentation framework that leverages contrastive learning to handle a large number of unlabeled samples. The core of this framework is the contrastive consistency regularization loss and class affinity loss, which together enhance the model’s ability to learn from unlabeled data. Our proposed contrastive consistency regularization loss incorporates pixel-level features from the entire batch into a cross-image consistency objective, substantially enhancing the model’s robustness to appearance variations. Complementarily, the class affinity loss leverages a compact memory bank of high-quality prototypes to directly refine the feature space—pulling unlabeled pixels toward their probable class centroids and pushing them away from alternative classes—thereby tightening intra-class compactness and sharpening inter-class boundaries for superior segmentation accuracy. The experimental results indicate that in the three different available labeled data experiments on NEU-SEG dataset, our proposed method improves the mIoU of the supervised learning paradigm by 4.4%, 3.7%, and 3.8%, respectively, surpassing the previous state-of-the-art method by 0.6%, 0.7%, and 0.5%. Finally, the ablation studies validate the effectiveness of the proposed method.