An Investigation into the Use of Vision Transformers for Industrial Anomaly Detection

Automated visual inspection is one of the most established applications of computer vision in industries and manufacturing. However, anomaly detection still remains an open challenge, as defects are usually rare, heterogeneous and expensive to collect in the industrially relevant regime, where defects’ examples are not always available at training time. This thesis aims to build on L2BT, a Teacher-Student method in which a frozen DINOv2 Teacher is imitated by two shallow MLP Students. In this approach, the Students learn to map the Teacher’s features from an early layer j to a deeper layer k, and wherever the mapping process fails an anomaly is localized. The method fixes the layer pair (j,k) by hand, and its authors acknowledge that this hyperparameter is under-explored. An exhaustive measurement of the layer-pair landscape on MVTec AD 2, covering all layer pairs on every category, confirms that no pair is universally optimal and that the cost of the wrong choice is large. Two label-free contributions are proposed in response. The first is a CKA fingerprint: the matrix of Centered Kernel Alignment between the first eight blocks of the Teacher (computed on fifteen nominal images) turns out to be a descriptor of the category rather than a performance predictor. This makes it possible to separate the categories into surface-like and object-like families. The family determines a small set of candidate pairs, which are then ranked on synthetic anomalies and combined into a z-normalized ensemble of two. The second is a peak-relative threshold, in which the binarization level is a fixed fraction of a quantile of the per-image response peaks. Evaluated on MVTec AD 2 under the official protocol, the pipeline improves the results of the original L2BT, while maintaining a low computational cost.