Embedding Stability-Guided Early-Stopping for Industrial Visual Anomaly Detection

Early-stopping is a widely used strategy to prevent overfitting in neural network training, but its standard forms rely on validation or loss signals that are often unavailable or unstable in industrial visual anomaly detection. This paper introduces a loss-independent early-stopping strategy based on the temporal consistency of learned embeddings, termed the stability divergence. The method detects convergence by monitoring embedding variations across epochs and terminates the model training once these variations fall below an adaptive threshold. The proposed approach is plug-and-play and operates without any access to test data. Experiments on the MVTec AD and VisA datasets demonstrate that the method achieves near-optimal performance on both GLASS and SimpleNet frameworks, offering a reliable and validation-free criterion for model checkpoint selection in industrial visual anomaly detection.