Abstract: A cascaded network model, GM-Net, is proposed to address the issues of unsupervised reconstruction networks trained on normal samples lacking defect information and insufficient ability to capture high-level semantic information during training. The model consists of a defect generator, a defect-guided reconstruction module, and a difference mapping siamese module. First, the defect generator is used to construct pseudo-defective samples that are unrelated to actual defects from normal samples, and corresponding normal and pseudo-defective features are extracted. Then, the defect-guided reconstruction module reconstructs randomly selected normal features or pseudo-defective features with normal features as the target. Finally, the difference mapping siamese module extracts the difference between the features before and after reconstruction and maps them to the image space to achieve defect detection. GM-Net enhances the representation ability of feature differences in the image space while guiding defect reconstruction. Performance tests were conducted on the KolektorSDD and KolektorSDD2 datasets, with AUC scores of 0.855 and 0.949, respectively.