EdgeZSAD: Practical Zero-Shot Anomaly Detection on Edge Devices
Researchers have developed EdgeZSAD, a practical system for zero-shot anomaly detection on edge devices, addressing the limitations of larger foundation models. The system utilizes a compact TinyViT-21M-512 backbone, an asymmetric global-local readout (EdgeGLR), and a reproducible training recipe (Real-IAD-DR). EdgeZSAD achieves strong performance on industrial benchmarks while being deployable on hardware like Jetson Orin Nano Super and RB5 Gen2, demonstrating minimal performance drift across different deployment settings. AI
IMPACT Enables more efficient and practical anomaly detection in industrial settings on resource-constrained edge devices.