N2AN-DA: a data augmentation pipeline for limited anomalous data using image-to-image translation

Industrial detection tasks frequently pay attention to anomalous data, yet their low occurrence probability leads to insufficient sample quantities. While generative data augmentation provides new solutions for the scarcity of anomalous data, practical implementations often concentrate on anomalous data alone, making it challenging to generate highquality samples under quantity constraints. We present the Normal-to-Anomalous Data Augmentation (N2AN-DA), a data augmentation framework leveraging image-to-image translation (I2IT) to transform plentiful normal samples into anomalous ones, thereby improving the volume and variety of anomalous instances. The method enhances detection models' robustness and adaptability, especially effective in environments with abundant normal data, scarce anomalies, and inherent connections between them. Our assessment using the liquid leakage trace dataset (W2LT) shows that with scarce anomalous data, samples generated by the N2AN-DA pipeline exhibit 80% lower Fréchet Inception Distance (FID) and 25% higher Inception Score (IS) than those directly produced by generative models. Detection models trained using augmented samples demonstrate over 10% increases in precision and mAP50 compared to non-augmented baselines, along with substantially improved generalization performance. The results demonstrate that the N2AN-DA's transformation of normal samples into premium-quality anomalies significantly boosts the performance of models in anomaly detection tasks.