Abstract: Aiming at the problems of large differences in diamond blade defect features and limited defect samples,an anomaly detection algorithm for diamond blade based on the improved U-Transformer feature reconstruction model was proposed.This method only needs to be trained with normal samples to complete the detection and localization of abnormal areas. Firstly,a frozen pre-trained deep Convolutional Neural Networks (CNN) model was used to extract multi-scale fusion features,amplifying the differences between normal and abnormal images. Then,the encoder-decoder feature reconstruction model based on the U-Transformer structure was constructed,and the feature similarity between the reconstructed features and the input features was calculated to generate the similarity response map. Finally,to eliminate the noise responses in the normal region,the similarity response map was corrected to obtain the anomaly score map by estimating the anomaly scale factor using the Multi-Layer Perceptron (MLP) network. The experimental results demonstrate that the proposed method achieves an Image AUROC index of 0.989 and a Pixel AUROC index of 0.992 on the diamond blade defect dataset,meeting the requirements for diamond blade anomaly detection.

Key words: diamond blade, anomaly detection, U-Transformer model, pre-training, MLP network, feature reconstruction