Defect detection method and device based on training of few normal sample data sets

The invention discloses a defect detection method and device based on training of a small number of normal sample data sets. The defect detection method comprises the following steps: acquiring a plurality of high-resolution normal sample images to make the normal sample data sets; constructing an image feature and edge feature extraction network model; processing a normal sample image and inputting the processed normal sample image into the network model to train the feature extraction capability; performing feature extraction and splicing on the plurality of processed normal sample images by using the pre-trained network model; a multi-head self-attention module is used for carrying out feature fusion on a splicing result, and a normal sample low-rank feature matrix is obtained through multivariate Gaussian distribution calculation; processing a high-resolution product image, and then performing feature extraction, splicing and feature fusion through the network model to obtain a feature vector of the high-