Blind Localization and Clustering of Anomalies in Textures

TL;DR

This paper tackles blind anomaly clustering in textures by introducing a two-stage pipeline: blind anomaly localization (BAL) using an extension of FCA on VAE residuals, followed by contrastive learning to refine features for clustering. The BAL step leverages both local feature residuals and global normality information from a VAE trained on the input set, producing accurate anomaly maps. A lightweight contrastive head $H$ is trained with carefully constructed positive/negative pairs to tighten intra-type similarity and separate different anomaly types, enabling effective clustering with Ward linkage on $H(F_i)$. Across MVTec AD textures, MTD, and Leaves, the approach achieves state-of-the-art clustering performance, demonstrating robustness to anomaly contamination and the challenge of small, subtle anomalies in textures. The method holds practical potential for unsupervised discovery of anomaly types in industrial texture inspection and related domains.

Abstract

Anomaly detection and localization in images is a growing field in computer vision. In this area, a seemingly understudied problem is anomaly clustering, i.e., identifying and grouping different types of anomalies in a fully unsupervised manner. In this work, we propose a novel method for clustering anomalies in largely stationary images (textures) in a blind setting. That is, the input consists of normal and anomalous images without distinction and without labels. What contributes to the difficulty of the task is that anomalous regions are often small and may present only subtle changes in appearance, which can be easily overshadowed by the genuine variance in the texture. Moreover, each anomaly type may have a complex appearance distribution. We introduce a novel scheme for solving this task using a combination of blind anomaly localization and contrastive learning. By identifying the anomalous regions with high fidelity, we can restrict our focus to those regions of interest; then, contrastive learning is employed to increase the separability of different anomaly types and reduce the intra-class variation. Our experiments show that the proposed solution yields significantly better results compared to prior work, setting a new state of the art. Project page: https://reality.tf.fau.de/pub/ardelean2024blind.html.

Figures (8)

• Figure 1: System overview. We train a VAE on features extracted from the input set, and apply FCA ardelean2024high on the residuals to obtain anomaly maps in a blind setting. These maps are used to mine positive and negative pairs for contrastive learning. The resulting improved image-level descriptors are then hierarchically clustered.
• Figure 2: Qualitative evaluation of blind localization and clustering of anomalies. Note that we only show a subset of each cluster; the proportion of misclassifications reflects the true accuracy (rounded down to include more failure cases). The color of a column name indicates the contour color associated with that anomaly type; green is reserved for the normal class (no anomaly). The contours for 'Ours' and '(Ours) w/o CL' differ only through their color because they start from the same binary anomaly localization (our BAL).
• Figure 3: Qualitative comparison of anomaly maps produced by different approaches to BAL. The images' borders are cropped out.
• Figure 4: Visualization of the initial image-level descriptors compared to the descriptors computed after contrastive learning. The vectors are projected to two dimensions using t-SNE.
• Figure 5: Analysis of clustering performance in terms of NMI for varying values of $\tau$. Contrastive learning improves stability.