Self‐Supervised Anomaly Segmentation for Surface Defect Inspection in Display Panels

Accurate segmentation of surface defects is essential for automated inspection in modern display panel manufacturing. Although recent studies have primarily focused on image‐level anomaly detection, pixel‐level anomaly segmentation remains underexplored due to the lack of annotated defect data. In this paper, we propose AnoSeg , a novel anomaly segmentation framework designed to directly generate accurate anomaly maps without requiring real defective samples. AnoSeg integrates three key techniques: (1) a hard augmentation strategy for synthesizing defect‐like anomalies from normal images, (2) self‐supervised learning with combined pixel‐wise and adversarial losses to enhance segmentation quality, and (3) coordinate channel concatenation to incorporate spatial priors relevant to defect localization in structured panel layouts. By training on only normal and synthetically augmented data, AnoSeg learns to segment anomalous regions robustly and generalizes well to unseen defects. Additionally, the resulting anomaly maps can be leveraged to improve conventional anomaly detection performance. Experimental results on the MVTec Anomaly Detection (MVTec‐AD) benchmark demonstrate that AnoSeg outperforms existing state‐of‐the‐art methods in both segmentation accuracy and detection performance, as measured by intersection over union (IoU) and AUROC metrics. These results suggest the practical applicability of AnoSeg for high‐precision surface inspection in intelligent display manufacturing systems.