Industrial Product Surface Defect Detection via Memory-Bank Feature Matching and Lightweight Diffusion Reconstruction

Surface defect detection of industrial products is a critical component of intelligent manufacturing quality control. However, existing unsupervised anomaly detection methods still struggle to achieve a favorable balance among image-level detection accuracy, pixel-level localization performance, and inference efficiency. To address this issue, this paper proposes a Fusion PatchCore with Lightweight Diffusion Reconstruction (FPC-LDR) framework for industrial surface defect detection, exploiting the complementary properties of discriminative feature matching and generative reconstruction. The proposed method first constructs a memory-bank module based on deep semantic features to identify anomalies via feature matching. Then, a lightweight latent diffusion reconstruction module is developed, in which single-step reconstruction replaces conventional multi-step diffusion processes, reducing computational overhead while maintaining reconstruction quality. Furthermore, an adaptive fusion strategy is introduced to integrate feature distance and reconstruction residuals, improving detection robustness and localization accuracy. Extensive experiments on the MVTec AD dataset demonstrate that FPC-LDR achieves an image-level AUROC of 94.65%, a pixel-level AUROC of 94.54%, and a PRO score of 95.28%. Cross-dataset evaluations on the VisA dataset further verify its generalization capability. The proposed method achieves an effective trade-off among accuracy, localization performance, and inference efficiency, demonstrating its potential for real-time industrial defect inspection.