A Prototype-oriented Fast Refinement Model for Few-shot Industrial Anomaly Detection
Keywords: Industrial Anomaly Detection, Few-shot Learning, Optimal Transport
Abstract: Industrial Anomaly Detection (IAD) in low data regime is crucial for automating industrial inspections in practice. Previous methods have primarily focused on obtaining robust prototypes using only a few normal images per product. However, these methods seldom account for transferring the characteristics of online query images to enhance the representativeness of the original prototypes in a systematic way. To address the pivot issue, we propose a fast prototype-oriented refinement model for few-shot IAD. Given online query images, we formulate prototype refinement as a nested optimization problem between transport probability for anomaly suppression and transform matrix for characteristic transfer. Then we present an Expectation Maximization (EM)-based algorithm to iteratively compute the transport probability and transform matrix. In the E-step, we use entropy-based optimal transport, known as the Sinkhorn algorithm, to learn the transport probability. In the M-step, the transform matrix is updated via gradient descent. Finally, we integrate our model with two popular and recently proposed few-shot IAD methods, PatchCore and WinCLIP. Comprehensive experiments on three widely used datasets including MVTec, ViSA, and MPDD verify the effectiveness and efficiency of our proposed model in few-shot IAD applications.
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.
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: 5785
Loading