WAAR-Net: Weakly-significant Anomaly Area Recognition Network for Visual Detection under Subtle Feature Conditions

In real-world visual scenarios, anomaly regions often exhibit weakly significant features that are difficult to detect due to low contrast and ambiguous boundaries. To address this challenge, we propose WAAR-Net, a novel network for visual anomaly area recognition under subtle feature conditions. WAAR-Net integrates multi-scale feature enhancement to amplify latent anomaly signals, contextual consistency learning to ensure coherent detection across local and global regions, and a learnable weak-feature attention mechanism to highlight anomalies while suppressing background noise. Extensive experiments demonstrate that WAAR-Net achieves superior detection accuracy and lower false positive rates compared with existing methods, particularly in scenarios with subtle or faint anomaly cues, making it suitable for applications in industrial inspection, surveillance, and medical imaging.