Title: Enhancing Robustness in Multi-Scenario Metal Surface Defect Detection via Domain-Specific Enhancements and a Contrastive Augmented Transformer

Abstract: Preserving product quality in industrial manufacturing hinges on effective metal surface defect detection. Nevertheless, this task is hindered by several obstacles, such as the scarcity of annotated data, the complexity of spotting subtle, multi-scale anomalies, and inadequate generalization capabilities across varied environments. To overcome these limitations, this study introduces a novel framework known as the Contrastive Augmented Transformer (CAT) for resilient defect identification. The CAT architecture utilizes a hierarchical Swin Transformer backbone and features a redesigned feature pyramid network, which effectively integrates low-level textural details with high-level semantic information. This design allows for the accurate modeling of subtle and multi-scale defect patterns. To improve performance under real-world noisy conditions, we introduce a domain-specific droplet augmentation technique. Additionally, a hard negative mining strategy is integrated into the contrastive loss function, thereby boosting the model’s discriminative power in regions where defects are ambiguous. Evaluations on the KolektorSDD2 dataset reveal that CAT attains a pixel-level AUROC of 99.54%, surpassing current state-of-the-art methods. Moreover, CAT demonstrates exceptional generalization and robustness when tested on three unseen datasets: KSDD1, MTD (for tile defects), and MSDD (for rail surface defects). These results highlight the framework’s potential for broad industrial application.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC