基于Swin Transformer和归一化流的色织物表面缺陷检测OA

Traditional deep learning methods are limited by the scarcity of defect samples,complex background and difficult identification of small target defects in fabric defect detection.In response to the problems,an unsupervised fabric defect detection and location method based on Swin Trans-former and normalized flow was proposed.First,in the training stage,only defect-free fabric im-ages were used to construct the training set,and Swin Transformer was used to extract multi-scale features.Then,a probability density estimation model was established using the normalized flow to model the distribution of normal sample features,so that the model can learn the potential spatial representation of normal fabric features.In the inference stage,the features of the fabric image to be measured were projected onto the learned feature distribution and their anomaly scores were calculated.Finally,the defect area of fabric was detected and located by anomaly fraction diagram.The experimental results show that this method can effectively learn the feature distribution of nor-mal fabrics and accurately detect and locate various fabric defects under complex background.On the YDFI-1 data set,the proposed method achieves 98.4％image-level AUROC and 96.9％pixel-level AUROC,which is significantly better than the existing unsupervised fabric defect detection methods.This method does not need defect samples and defect labeling,and only relies on the fea-ture distribution of normal samples for defect identification,thus improving the generalization ability and robustness of detection.