ViLoc-Net: Leveraging Synthetic Defect Generation and Vision Transformers for Industrial Surface and Texture Anomaly Detection

초록

Automated visual inspection in industrial settings often struggles with limited defect data and poor generalization to unseen anomalies. To overcome this challenge, we propose a hybrid anomaly detection pipeline, which integrates embedding-based, reconstruction-based, and self-supervised learning approaches. The framework also proposes a new Realistic Industrial Defect Synthesis (RIDS) module that synthesizes structured and textured synthetic anomalies based on the target masks, composite maps, and blending techniques. This helps to learn from pseudo-labeled data without the need for large annotated datasets. The pipeline further includes ViLoc-Net, a Vision Transformer-based localization network that obtains global features and then reconstructs detailed segmentation maps by a multi-scale decoder. The method is tested on various industrial datasets, demonstrating high accuracy, robustness, and generalization, making it suitable for real-world inspection tasks.

키워드