SYNAD: A synthetic object injection methodology for enhanced anomaly detection

Abstract

In vision-based anomaly detection, the scarcity of anomalous samples in real-world datasets is one of the main challenges. If none or only a few labeled anomalies are available, this poses a challenging constraint to developers and significantly limits the design space for an anomaly detection system. In this paper, we introduce a methodology for inserting synthetic anomalies into images to extend the anomaly detection design space for real-world datasets where no anomaly labels are available for training. We focus on the public railway dataset “Nordland” and demonstrate how our synthetic anomaly insertion methodology enables the training of classification and segmentation networks and can outperform state-of-the-art reconstruction-based approaches. We annotated real anomalies in sections of the “Nordland” dataset for analysis to demonstrate the effectiveness of our anomaly detection training methodologies. Our proposed segmentation training methodology increases the baseline model's recall rate by up to seven percentage points at the same false-positive rate while eliminating the need for a second network. Furthermore, our binary classification approaches increase the recall rate by up to 28 percentage points while reducing the false positive rate of the baseline from 5% to 3.47%.