Original title: PL-SCEA: Reconfiguring Pretrained Attention for Few-Shot Industrial Anomaly Detection
Xiaoyu Yang, Qixing Wu, Huixian Zhao, Changlong Jin · September 2026
Factory cameras need to spot tiny defects in products using only a few normal reference pictures. Computer vision models are good at recognizing general objects, but they often miss small texture flaws because their internal attention focuses on high-level meaning instead of tiny physical details. To fix this mismatch, researchers developed Power-Law Self-Correlation Enhanced Attention, or PL-SCEA. The method alters the attention inside a frozen vision model without adding trainable projection layers. It keeps the model's original context but calculates new correlations between image tokens, filtering for positive links and applying a power-law weighting to highlight unusual patterns against their background. Next, a lightweight variational autoencoder learns what normal products look like and scores defects by measuring reconstruction differences. Tested on MVTec AD and VisA, the system delivered competitive defect detection and consistently strong pixel-level flaw localization across few-shot settings.
When you walk into a familiar room, your eyes take in the whole scene automatically. But if something is subtly off, like a slight change in how objects are arranged, you might miss it unless you're specifically looking for relational details. In this paper, the AI's attention is adjusted to focus more on how different parts of an image relate to each other in texture and structure, making it better at catching those subtle anomalies.
In factories, catching defects early saves money and prevents failures. This method lets AI learn from just a few examples, making quality control more efficient without needing massive datasets.
The abstract doesn't specify if this method works for all defect types or industries. It is evaluated only on specific datasets (MVTec AD and VisA), so performance elsewhere is not established. The approach depends on frozen pretrained models, so effectiveness may vary with different base models.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
PL-SCEA: Reconfiguring Pretrained Attention for Few-Shot Industrial Anomaly Detection — Xiaoyu Yang, Qixing Wu, Huixian Zhao et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.