Forward-Backward Feature Transfer for Industrial Anomaly Detection and Segmentation
Keywords: anomaly, detection, segmentation, localization
TL;DR: A novel fast method for industrial anomaly detection and segmentation.
Abstract: Motivated by efficiency requirements, most industrial anomaly detection and segmentation (IADS) methods process low-resolution images, e.g., $224\times 224$ pixels, obtained by downsampling the original input images.
In this setting, downsampling is typically applied also to the provided ground-truth defect masks.
Yet, as numerous industrial applications demand the identification of both large and small defects, this downsampling procedure may fail to reflect the actual performance achievable by current methods.
In this work, we propose a fast approach based on a novel Teacher-Student paradigm.
This paradigm relies on two shallow Student MLPs that learn to transfer patch features across the layers of a frozen Teacher Vision Transformer.
Our framework can spot anomalies from high-resolution images faster than other methods, even when they process low-resolution images, achieving state-of-the-art overall performance on MVTec AD and segmentation results on VisA.
We also propose novel evaluation metrics that capture robustness regarding defect size, i.e., the ability of a method to preserve good localization from large anomalies to tiny ones, focusing on segmentation performance as a function of anomaly size.
Evaluating our method with these metrics reveals its stable performance in detecting anomalies of any size.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7587
Loading