ObjSegAD-Net: Region-aware pseudo-defect injection and dual-branch architecture for unsupervised industrial anomaly detection

Abstract

Anomaly detection in industrial manufacturing identifies product defects. However, limited anomalous samples and background noise often reduce accuracy and increase false alarms. To solve this, we propose ObjSegAD-Net, an unsupervised anomaly detection method. It separates each image into foreground and background regions. It uses foreground-guided pseudo-defect synthesis to inject diverse synthetic anomalies, boosting data diversity with more pseudo-samples. For the background, it adds noise such as Gaussian blur, helping the model distinguish true defects from irrelevant variations. During inference, ObjSegAD-Net adopts a dual-branch architecture with region-aware attention mechanisms, which adaptively enhances responses to foreground anomalies while suppressing background interference. This design significantly reduces false positives and improves detection accuracy under complex noisy conditions. ObjSegAD-Net achieves state-of-the-art results on multiple industrial anomaly detection benchmarks, demonstrating the robustness and generalization capabilities of its region-aware pseudo-defect generation and dual-task architecture.