ViGLAD: Vision graph neural networks for logical anomaly detection

Anomaly detection is a research field that has been growing in the recent years with emerging industrial interest shown by the publication of multiple anomaly detection image datasets with a rising interest in logical anomaly detection. The challenge of logical anomaly detection lies in the nature of these

anomalies which, in contrast to structural anomalies, are hidden in the global relations between the image components. This work proposes using the graph representation of an image in order to tackle this challenge by proposing a novel approach called Vision Graph based Logical Anomaly Detection (ViGLAD). Defining

an image as a structure of nodes and edges leverages new possibilities for detecting logical anomalies by introducing vision graph auto-encoders. Our experiments on public datasets show that using vision graphs enhances the performance of state-of-the-art teacher-student-auto-encoder neural networks in logical anomaly detection while keeping a robust performance in structural anomaly detection.