An efficient anomaly detection method for the industrial sewing process using PatchCore and fine-tuned ResNet

Abstract In the fashion industry, sewing defects have traditionally been detected through manual visual inspection, which is limited in accuracy and speed. This study developed an efficient anomaly detection method for the industrial sewing process using PatchCore, an unsupervised feature-embedding model. The approach fine-tunes ResNet submodels of varying depths (18, 34, 50, 101, and 152) on the StitchingNet dataset to extract sewing-specific patch features, enhancing detection performance while using computing resources efficiently in small-scale manufacturing. The proposed model was compared with other unsupervised methods (the original PatchCore, PaDiM, Deep SVDD, and GANomaly), and the effect of submodel depth, layer combinations, and subsampling ratios was evaluated to enable rapid detection with limited data. On nylon-fabric images, it achieved an AUROC of 0.988, improving on the original PatchCore (0.862) through backbone fine-tuning and optimized configuration. Across all eleven StitchingNet fabric types, it yielded a mean AUROC of 0.984 and a mean F1-score of 0.983. An F1-score of 0.950 and an inference time of 0.37 s per image on a low-cost Raspberry Pi 5 confirm its suitability for deployment. Thus, the method suits sewing processes with limited training data and can serve as a core anomaly detection technology for clothing manufacturing.