DetectGPT：无标签特征互打分机制的无监督工业异常检测方法

打开文本图片集

中图分类号：TB9;TP391.41 文献标志码：A 文章编号：1674-5124（2025）09-0024-11

Abstract： Aiming at the problem that unsupervised industrial anomaly detection model needs to manually set theanomaly score threshold to distinguish between normal and abnormal samples，the detection process of multiple rounds of dialog and the difficulty of describing the attribute information of abnormal regions， this paper proposes an unsupervised large-scale industrial anomaly detection method DetectGPT， which constructs a training dataset based on the simulated anomalous images of industrial domain knowledge and generates the textual descriptions of each image containing the knowledge of the industrial domain.Using the implicit normal a priori information in the unlabeled images to asist in the detection of anomalies，the attribute information prompts the learner to fine-tune the large-scale language model to realize multi-round interactive anomaly detection，and outputs normal and abnormal region atribute information. Tests on MVTec-AD and

VisA datasets show that DetectGPT has good context learning ability in unsupervised less sample seting， ImageAUROC Σ=97.0% andPixelAUROC Ψ=93.6% onMVTec-ADdataset，and providesdetailed normal and abnormal image attribute information description， which has the prospect of wide application. Keywords： industrial anomaly detection; unsupervised; unlabeled; mutual scoring

0 引言

工业异常检测是生产质量控制重要环节[1-2]，广泛应用于零件加工[3-5]、使役过程[6-8]、产品装配[9-11]等场景中。（剩余15892字）