Zero Shot Anomaly

Zero-shot anomaly detection (ZSAD) aims to identify anomalies in unseen data without requiring training on examples of the target anomaly type. Current research heavily utilizes large vision-language models (like CLIP) and large language models (LLMs), often incorporating techniques such as learnable prompts, multi-modal fusion, and visual context prompting to improve anomaly detection and localization accuracy. This field is significant because it addresses the limitations of traditional anomaly detection methods that require extensive labeled data, enabling applications in diverse areas such as industrial quality control, medical image analysis, and cybersecurity where obtaining sufficient labeled data is challenging or impossible.