A Defect-Inspection System Constructed by Applying Autoencoder with Clustered Latent Vectors and Multi-Thresholding Classification

Defect inspection is an important issue in the field of industrial automation. In general, defect-inspection methods can be categorized into supervised and unsupervised methods. When learning applied to defect inspection, large variation patterns make data coverage incomplete for model training, which introduce problem low detection accuracy. Therefore, this paper focuses on construction a system with model. Furthermore, few studies have focused analysis between reconstruction error normal areas repair effect defective systems. Hence, addresses issue. There are four main contributions paper. First, we compare effects SSIM (Structural Similarity Index Measure) MSE (Mean Square Error) functions error. Second, various kinds Autoencoders constructed by referring Inception architecture GoogleNet DEC (Deep Embedded Clustering) module. Third, two-stage training proposed train Autoencoder models. first stage, models trained basic image-reconstruction capabilities areas. second algorithm added further strengthen feature discrimination then increase capability Fourth, multi-thresholding image segmentation method improve classification accuracy images. study, focus texture patterns. select nanofiber database carpet grid images MVTec conduct experiments. The experimental results show that classifying patch about 86% approach 89% 98% datasets database, respectively. It obvious our outperforms MVTec.