HSS-IAD: A Heterogeneous Same-Sort Industrial Anomaly Detection Dataset

TL;DR

This work identifies a gap between real factory anomaly detection needs and existing MUAD datasets, notably the lack of same-sort intra-factory variation and subtle defects. It introduces HSS-IAD, a large, heterogeneous dataset of 8,580 metal- or magnetic-tile images across 7 categories, with precise defect annotations and foregrounds for synthetic anomaly generation, plus two evaluation settings. The authors benchmark current SoTA MUAD methods on HSS-IAD, showing the dataset is more challenging than existing benchmarks and highlighting substantial room for improvement in both detection and localization under realistic industrial conditions. Overall, HSS-IAD aims to bridge the gap between academic datasets and practical factory scenarios, driving the development of more robust unified anomaly detection techniques for industrial surfaces.

Abstract

Multi-class Unsupervised Anomaly Detection algorithms (MUAD) are receiving increasing attention due to their relatively low deployment costs and improved training efficiency. However, the real-world effectiveness of MUAD methods is questioned due to limitations in current Industrial Anomaly Detection (IAD) datasets. These datasets contain numerous classes that are unlikely to be produced by the same factory and fail to cover multiple structures or appearances. Additionally, the defects do not reflect real-world characteristics. Therefore, we introduce the Heterogeneous Same-Sort Industrial Anomaly Detection (HSS-IAD) dataset, which contains 8,580 images of metallic-like industrial parts and precise anomaly annotations. These parts exhibit variations in structure and appearance, with subtle defects that closely resemble the base materials. We also provide foreground images for synthetic anomaly generation. Finally, we evaluate popular IAD methods on this dataset under multi-class and class-separated settings, demonstrating its potential to bridge the gap between existing datasets and real factory conditions. The dataset is available at https://github.com/Qiqigeww/HSS-IAD-Dataset.

Figures (6)

• Figure 1: Comparison of training paradigms.
• Figure 2: Normal and anomalous samples from existing datasets and the HSS-IAD dataset. Close-ups of anomalies are highlighted in red boxes. Compared to existing datasets (e.g., MVTecAD, RealIAD), which include diverse classes like fruits, snacks, pills, etc., the HSS-IAD dataset has the following characteristics: 1.) The subclasses consist of same sort industrial products while exhibiting variations in structure or appearance. 2.) The subtle and variable defects closely resemble the material itself.
• Figure 3: Data collection pipeline for our proposed HSS-IAD dataset. (a) The collection procedure encompasses various stages, including data collection, deduplication, foreground generation, label generation, label refinement, and reclassification and filtering. (b) The samples in the HSS-IAD dataset are divided into the training set and the test set. The training set consists solely of normal samples of industrial products made of metal or magnetic tile. These products have same sort but differ significantly in structure or appearance. The test set consists of normal samples and abnormal samples with subtle and variable defects. Close-up figures of abnormal samples, with highlighted abnormal areas, are displayed in the last row along with their pixel-precise ground truth labels.
• Figure 4: Statistical information of our proposed HSS-IAD dataset: (a) Distribution of anomaly/normal image quantities in training and test across different categories. (b) Statistics of the percentage of the image area occupied by anomaly region. (c) Statistics of the aspect ratio of the minimum bounding rectangles of the defect.
• Figure 5: (a) Casting image; (b) Foreground image of Casting.