Patchcore Transformed: A Vision Transformer Approach To Anomaly Detection

Recent advancements in natural language processing (NLP) have spurred innovation in computer vision, leading to the introduction of Vision Transformers (ViT). These transformer-based models have demonstrated remarkable performance, surpassing traditional Convolutional Neural Networks (CNN) in various computer vision tasks. This thesis explores the integration of ViT into PatchCore, an anomaly detection model, to leverage the capabilities of transformer architectures for anomaly detection tasks. Experiments are conducted on two datasets: MVTec AD, a widely used benchmark dataset, and a custom dataset tailored to specific application domains. The results demonstrate the effectiveness of the ViT-PatchCore model in detecting anomalies and outperforming existing methods, including traditional CNN-based approaches. Furthermore, this thesis proposes a novel solution that addresses some challenges in integrating ViT into PatchCore and suggests improvements to its algorithm to enhance its performance and efficiency. This research contributes to the advancement of anomaly detection techniques by harnessing the power of transformer based models in computer vision applications and suggests potential directions for future research in this field.