<p>Unsupervised anomaly detection in industrial images is a topic worthy of attention. In the task of anomaly detection, the teacher–student (T–S) model based on knowledge distillation has been proven to be reasonable and effective. However, traditional T–S-based anomaly detection algorithms are only trained on normal samples, and the student network lacked the ability to restore abnormal structures in images, which affected the final detection effect. Therefore, this paper proposed a defect detection and location method based on adaptive multi-branch residual denoising and triple squeeze excited attention mechanism (ADTE). Firstly, four parallel multi-branch residual blocks are embedded into the student network and set to adaptive mode to remove abnormal features in the image. Secondly, to enhance the model’s focus on important features, an improved triple squeeze excited attention mechanism is proposed to reassign new weights to the output features, further strengthening the student network’s ability to remove anomalies. Finally, to enable the network to have the ability to recognize anomalies, an irregular anomaly synthesis method is proposed, and the synthesized abnormal images and corresponding normal images are respectively input into the T–S network for training. After a substantial amount of experiments, it is shown that our model achieves very effective results on the large industrial datasets MVTecAD and VisA. The average image level AUC-ROC, pixel level AUC-ROC, AUC-PRO on MVTec AD are 98%, 98% and 94.1% respectively. The average pixel-wise AUC-ROC on VisA is 98.5%.</p>