The invention discloses an industrial
anomaly detection method based on unmarked model search, and the method comprises the steps: searching and selecting a plurality of vision-language models in an
open source model
community to construct a pre-training model
library, and collecting a plurality of sample images of a to-be-detected industrial element to form a
training set; constructing an exception prompt text, inputting the exception prompt text into an
open source large
language model to generate a plurality of exception types and constructing an exception type text description, and performing zero sample prediction on the
sample image and the exception type text description based on a vision-
language model to obtain a pseudo-fact prediction result; and calculating the consistency between the prediction result of each vision-
language model and the pseudo-fact prediction result, carrying out sorting selection to obtain a reuse model, inputting the to-be-detected
sample image and the abnormal category text description into the reuse model, and respectively obtaining a final abnormal
score generated after detection output fusion. According to the method, effective industrial
anomaly detection can be realized without any manual
data annotation and large-
scale model training, and the method has the characteristics of high applicability, low cost and high precision.