Enforced clustering for zero-to-one-shot texture anomaly detection

Recent studies on anomaly detection (AD) for industrial products typically address the problem in an unsupervised manner, requiring only normal data for training. This approach alleviates the need for anomalous data but still requires a set of normal samples and often involves demanding computations. More recent methods aim to solve this problem in zero-, one-, or few-shot settings but suffer from performance drops or rely on additional contexts, such as language guidance and text encoding, which add overhead. This paper focuses on homogeneous textures and demonstrates how the problem can be addressed without any training samples or additional training (zero-shot), only requiring one normal sample (one-shot) for hyperparameter selection, which is an additional challenge in unsupervised settings. This is achieved by enforcing K-means clustering with on each of the testing samples independently, distinguishing it from the typical use of clustering methods in outlier detection, which are applied to a set of samples. The confidence score of each locality belonging to the smaller cluster, considered the potential anomalous cluster for evaluation, forms the anomaly map used in anomaly localization, and the maximum values in this map are used in AD. Competitive performance is achieved through the careful selection of the distance metric, feature layers, and clustering method. Experiments show that this zero-to-one-shot method, which facilitates deployment by reducing data dependency, maintains performance comparable to, or even higher than, conventional many-shot methods, all with relatively high speed.