ViT-Core: Lightweight Anomaly Detection Model using Transformer-based Feature Extractor

In modern smart manufacturing, real-time anomaly detection is critical, yet state of the art (SOTA) models like PatchCore are often too computationally demanding for practical deployment on resource -constrained edge devices. This creates a crucial gap between algorithmic potential and industrial application. To bridge this gap, we propose ViT-Core, a lightweight and efficient model for anomaly detection. Our primary innovation involves replacing the conventional CNN backbone with an efficient Swin Transformer, which reduces feature map dimensionality and, consequently, memory usage. To maintain high accuracy, we employ a Cut-Paste-based transfer learning stage, a self-supervised process that fine-tunes the model to the target data distribution without requiring complex training or additional labels. Evaluated on the comprehensive MVTec AD benchmark, ViT-Core demonstrates a drastic reduction in computational overhead, with memory usage decreased by 49.2% and inference time by 49.5% compared to the PatchCore baseline. This optimization is achieved with a statistically insignificant difference in image-level classification performance (0.9859 AUROC vs. 0.9865). Moreover, ViT-Core excels in anomaly localization, improving the pixel-level AUROC to 0.9817 from PatchCore's 0.9756. Consequently, ViT-Core presents an optimal balance of accuracy and efficiency, providing a practical and scalable solution that enables the widespread deployment of high-performance, real-time quality inspection systems directly on existing industrial hardware.