-Title: REFRACT: Restoration-Adapted Reconstructive-Discriminative Anomaly Detection for Industrial Inspection
-Journal/Conference: ACCV 2026
-Authors:  Jaehoon Lim, Yeonkyu Kwak and Jongpil Jeong
-DOI: 
-Journal/Conference Link: https://accv2026.org/

Abstract: Industrial anomaly detection, which automatically identifies surface defects on manufactured products, is a key technology for quality control, yet defective samples and pixel-level annotations are scarce. Conventional reconstruction-based methods can be trained on normal images alone, but relying directly on reconstruction error causes defects to be reconstructed together with the input, or normal-region errors to be mistaken for anomalies. We propose REFRACT, a reconstructive–discriminative framework that adapts a frozen, pretrained image restoration model for anomaly detection rather than training a reconstructor from scratch. A LoRA adapter keeps only about 4.26% of the backbone trainable, while a lightweight discriminative U-Net takes a 9-channel input (the input image, the reconstruction, and their absolute difference) to predict a pixel-level anomaly map, decoupling detection from reconstruction fidelity. Anomalies are synthesized from normal images by blending DTD textures under a curriculum schedule. Instantiated with the MambaIRv2 restoration backbone, REFRACT attains a mean image-level AUROC of 0.8272, a pixel-level AUROC of 0.8787, and a PRO score of 0.7095 on the 15 categories of MVTec AD. This mean conceals a sharp split: REFRACT is competitive on texture categories (0.958 image-level AUROC) but weak on object categories (0.762), positioning it as a texture-oriented detector.
GitHub - https://github.com/jhlimdev

-Status: Submitted (2026/06/27)