FUN-AD: Fully Unsupervised Learning for Anomaly Detection with Noisy Training Data

TL;DR

This work proposes a novel learning-based approach for fully unsupervised anomaly detection with unlabeled and potentially contaminated training data using an iteratively reconstructed memory bank (IRMB), and proposes a pseudo-labeling strategy using an iteratively reconstructed memory bank (IRMB).

Abstract

While the mainstream research in anomaly detection has mainly followed the one-class classification, practical industrial environments often incur noisy training data due to annotation errors or lack of labels for new or refurbished products. To address these issues, we propose a novel learning-based approach for fully unsupervised anomaly detection with unlabeled and potentially contaminated training data. Our method is motivated by two observations, that i) the pairwise feature distances between the normal samples are on average likely to be smaller than those between the anomaly samples or heterogeneous samples and ii) pairs of features mutually closest to each other are likely to be homogeneous pairs, which hold if the normal data has smaller variance than the anomaly data. Building on the first observation that nearest-neighbor distances can distinguish between confident normal samples and anomalies, we propose a pseudo-labeling strategy using an iteratively reconstructed memory bank (IRMB). The second observation is utilized as a new loss function to promote class-homogeneity between mutually closest pairs thereby reducing the ill-posedness of the task. Experimental results on two public industrial anomaly benchmarks and semantic anomaly examples validate the effectiveness of FUN-AD across different scenarios and anomaly-to-normal ratios. Our code is available at https://github.com/HY-Vision-Lab/FUNAD.

Figures (8)

• Figure 1: An illustration of our motivation. We assume that the (image-level and patch-level) features from the normal samples (in green) exhibit smaller variance than those from the anomaly samples (in red). In Sec. \ref{['sec:statistical_analysis_pairwise_dist']}, we analytically show this leads to a homogeneous pair of normal samples being more likely to yield smaller pairwise distance than other types of pairs. In Sec. \ref{['sec:smoothness_motivation']}, we empirically show that mutually closest pairs are likely to be homogeneous pairs (i.e. mostly normal-normal or anomaly-anomaly).
• Figure 1: Histogram of pairwise distances for different types of feature pairs. Abbreviations are as follows: NN for normal-normal pairs, AA for anomaly-anomaly pairs, AN or NA for anomaly-normal or normal-anomaly pairs. For the synthetic experiment, the normal samples were drawn from $\mathcal{N}(\mathbf{0}, \mathtt{I})$ and the anomaly samples from $\mathcal{N}(\mathbf{1.5}, 2 \mathtt{I})$. For the "real" comparison, we used patch features (patch tokens) and image features (class tokens) extracted from the pretrained DINO model caron2021emerging for the bottle set bergmann19mvtec.
• Figure 2: Visualization of the matching ratios for different types of feature pairs. The empirical experimental settings are as in Sec. 1 in supmat for both synthetic and real. For "true (normal)", the matching ratio is the number of samples mutually closest to normal-normal divided by the number of normal samples, and for "true (anomaly)", the matching ratio is the number of samples mutually closest to anomaly-anomaly divided by the total number of anomaly samples. Additionally, "false" represents the number of samples mutually closest to normal-anomaly (or anomaly-normal) divided by the total number of samples (including both normal and anomaly).
• Figure 2: Visualization of anomaly detection results achieved by FUN-AD on the MVTec AD and VisA datasets. Each binary mask shows the anomaly segmentation map while each heatmap visualizes the anomaly region (red means likely to be an anomaly while blue means unlikely).
• Figure 3: Our framework overview for fully unsupervised anomaly detection. While the framework itself is simple, its constituent components such as patch-level pseudo-label generator and mutual smoothness loss are designed to effectively utilize the observations in Sec. \ref{['sec:motivation']}.