Fractals as Pre-training Datasets for Anomaly Detection and Localization

ICLR 2024 Workshop DMLR Submission38 Authors

Keywords: anomaly detection, fractals images, data generation, feature-embedding

TL;DR: We use synthetically generated fractal images to pre-train models for the task of unsupervised industrial defect detection

Abstract: Anomaly detection is a crucial application in large-scale industrial manufacturing as it helps detect and localise defective parts. Pre-training feature extractors on large-scale datasets is a popular approach for this task. However, creating such large datasets is expensive and time-consuming and requires careful investigation of technical and social issues. While recent work in anomaly detection primarily focuses on the development of new methods built on such extractors, the importance of the data used for pre-training has not been studied. Therefore, we evaluated the performance of eight state-of-the-art anomaly detection methods pre-trained using dynamically generated fractal images on the famous benchmark datasets MVTec and VisA. In contrast to the literature that focused on fractals’ transfer-learning ability, in this study, we compared models pre-trained with fractals against ImageNet without fine-tuning. Although pre-training with ImageNet remains a clear winner, the results of fractals are promising considering that this task required features capable of discerning even minor visual variations and we can do that without fine-tuning the weights, thereby lacking familiarity with the dataset. This opens the possibility for a new research direction where feature extractors could be pre-trained with synthetically generated abstract datasets overcoming the problem of privacy, basis and inappropriate content, as no humans are pictured.

Primary Subject Area: Data-centric approaches to AI alignment

Paper Type: Research paper: up to 8 pages

DMLR For Good Track: Participate in DMLR for Good Track

Participation Mode: In-person

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Submission Number: 38