Supervised Anomaly Detection for Complex Industrial Images

TL;DR

This paper addresses the gap between public anomaly-detection datasets and real-world industrial inspection by introducing the Valeo Anomaly Dataset (VAD), a real-world, supervised AD benchmark with substantial defect diversity and unseen defect types. It proposes SegAD, a segmentation-informed anomaly detector that aggregates anomaly maps over predefined segments, computes local statistics, and uses a Boosted Random Forest to produce final anomaly scores, optionally incorporating a supervised classifier score. Across one-class, high-shot, and low-shot supervised regimes on VAD and on the VisA dataset, SegAD achieves state-of-the-art performance, demonstrating robustness to unseen defects and improved accuracy without requiring pixel-level defect masks. The work also provides a practical, scalable framework and release-ready code to bridge research and industrial deployment in visual defect detection.

Abstract

Automating visual inspection in industrial production lines is essential for increasing product quality across various industries. Anomaly detection (AD) methods serve as robust tools for this purpose. However, existing public datasets primarily consist of images without anomalies, limiting the practical application of AD methods in production settings. To address this challenge, we present (1) the Valeo Anomaly Dataset (VAD), a novel real-world industrial dataset comprising 5000 images, including 2000 instances of challenging real defects across more than 20 subclasses. Acknowledging that traditional AD methods struggle with this dataset, we introduce (2) Segmentation-based Anomaly Detector (SegAD). First, SegAD leverages anomaly maps as well as segmentation maps to compute local statistics. Next, SegAD uses these statistics and an optional supervised classifier score as input features for a Boosted Random Forest (BRF) classifier, yielding the final anomaly score. Our SegAD achieves state-of-the-art performance on both VAD (+2.1% AUROC) and the VisA dataset (+0.4% AUROC). The code and the models are publicly available.

Figures (9)

• Figure 1: Overview of our contributions. (a) VAD, a real-world industrial dataset designed for supervised anomaly detection with complex defects. (b) SegAD, our method that leverages anomaly maps extracted from segmented outputs of one or more anomaly detectors. Higher-level statistical features are computed on these maps, such as skewness, kurtosis, or mean, to generate the local anomaly features. Additionally, our SegAD provides the flexibility to use the output of a supervised classifier score with the local anomaly features, creating input for a final Boosted Random Forest (BRF) classifier that yields the final score.
• Figure 2: Good parts. Small scratches on the piezo are allowed, as well as wire being closer to the side of the soldering. (VAD)
• Figure 3: Variety of defects: logical, from the top left: wire on the side, wire out of solder dot, missing wire, bad wire shape (bent too much). Structural: soldering paste pollution, crack on a piezo, burned area under the solder dot, broken piezo. (VAD)
• Figure 4: Unseen defects for left to right: burned solder dot, burned wire, extra unnecessary wire. (VAD)
• Figure 5: Static segmentation map for objects in VAD dataset, overlayed over the image. The image is separated into $L=7$ segments: background, outer half-circle, piezo border, solder dots area, wires area, pins area, and piezo in the middle.