Anomaly detection for real-world industrial applications: benchmarking recent self-supervised and pretrained methods

Visual anomaly detection (AD) is currently a very active research area with great potential in many real-world applications, e.g. quality control in industry and manufacturing, where it can provide cost savings and overall better product quality. Recently, many new anomaly detection methods have been introduced and many of them are intended for industrial usage. These methods are usually evaluated on a narrow selection of datasets that may differ significantly from certain types of real-world applications. Due to this approach, some methods provide different performance when deployed in real use cases. In this paper, we perform evaluation of recent state of the art visual anomaly detection methods on the problem of defect detection in metal parts fabrication, an area not well covered in existing publications. We introduce a new dataset focused specifically on the problem of metal parts fabrication and use the dataset to perform the evaluation. For the evaluation, we selected methods that use two different feature extraction approaches for anomaly detection. One of the approaches is using feature extractors pretrained on the ImageNet dataset and the second approach is training the feature extractor from scratch using self-supervised learning. We show that, in contrast to one of the most widely used anomaly detection benchmark - the MVTec-AD dataset, self-supervised methods perform significantly better (in average by 17 % AUROC) on the proposed dataset.