Abstract: To solve the problem of low feature reconstruction accuracy in industrial product surface defect detection, which leads to high false positive rates at the image, pixel, and region levels, we propose an unsupervised defect detection method with improved deep feature reconstruction (DFR). First, we introduce jump connections into the feature reconstruction process to improve the feature reconstruction accuracy and the ability of the model to reconstruct positive sample features. Second, we introduce an attention mechanism to improve the attention of the algorithm to defect regions and explore the effect of spatial attention on defect detection for different targets. Third, we introduce atrous pyramid pooling into the feature reconstruction module to capture context at multiple scales without increasing the number of parameters to improve the ability of the model to detect defects at different scales. Finally, we use the L₂-SSIM loss function to constrain feature reconstruction to preserve the feature structure while maintaining pixel similarity. The proposed algorithm achieves 97.5%, 97.2%, and 93.1% detection accuracy at the image, pixel, and region levels, respectively, outperforming the comparison algorithm.