Anomaly Detection in Industrial Inspection Visual Systems using Swin Transformer with Bayesian Optimization

In recent years, industrial visual inspection systems have become increasingly essential that recognize defects from normal patterns in images to ensure product quality and identify defective items in manufacturing systems. However, the existing Vision Transformer (ViT) failed to capture hierarchical feature representations and also faced high computation complexity. Therefore, to address these limitations, a Swin Transformer with Bayesian Optimization (Swin-Bo) is proposed for anomaly detection in industrial vision inspection system. Initially, the input data which consists of industrial defect images is collected from MVTec Anomaly Detection (AD) dataset. Next, these images are preprocessed using resizing to ensure standardize dimensions and normalized using min max normalization. Finally, these normalized images are fed to the Swin Transformer, which uses window-based attention mechanism for extraction of both local and global features. Therefore, the proposed Swin-Bo obtained better results in terms of Area Under the Receiver Operating Characteristic Curve (AUROC) and Per Region Overlap (PRO) by outperforming the existing ViT model.