EFFD: An Unsupervised Surface Defect Detection Method Based on Estimation and Fusion of Normal Sample Feature Distribution

Deep feature reconstruction (DFR) as an unsupervised surface defect detection framework, first combined the idea of feature embedding with reconstruction. However, it concatenated all features for reconstruction, which may cause the detection results to be disturbed by some features that are not sensitive to anomalies, causing missed detection. Therefore, we made a series of improvements to DFR and proposed a method called estimation and fusion of normal sample feature distribution (EFFD), as well as its upgraded version EFFD+. Our improvements are based on the assumption that features closer to the normal sample feature distribution are more sensitive to anomalies. We used multivariate Gaussian distribution estimation (MGDE) and Bayesian multivariate Gaussian distribution fusion (BMGDF) modules to estimate and fuse the normal sample feature distribution across all levels at each scale, and made the features fused by iterative attention feature fusion (IAFF) module follow to the fused distribution of the corresponding scale. This allowed the parts of the features that are sensitive to anomalies to be assigned higher weights. In EFFD+, we proposed IAFF+, which added more branches of different receptive fields based on IAFF and used weighted multiscale feature fusion (WMSFF) module to assign the optimal weights to the features at each scale. EFFD+ is more conducive to the detection of various categories of defects. On MVTecAD, BTAD, and VisA datasets, our proposed method achieved 98.5%, 97.9%, and 98.8% average pixel-level area under the receiver operating characteristic curve (P-AUC-ROC), respectively, and 94.6%, 81.9%, and 93.1% average area under the per region overlap curve (AUC-PRO), respectively, which has reached the current advanced level. Code is available at: https://github.com/Jalexdalv/effd.