LiZAD: A Lightweight Zero-Shot Anomaly Detection Framework for Industrial Manufacturing

2026-07-02

AI Summary

This work addresses the challenge of zero-shot anomaly detection in industrial manufacturing, where frequent changes in product configuration and appearance hinder defect data collection, and existing methods incur prohibitive computational costs for edge deployment. The study proposes a novel approach that integrates the high-accuracy vision encoder DINOv3 with the lightweight multimodal model MobileCLIP2, connected via a low-memory, trainable projection head to map features into a shared latent space. This design enables efficient zero-shot anomaly detection while preserving strong performance—exhibiting only a 6.4% drop in P-AUROC—yet significantly reduces resource demands: memory usage and parameter count are lowered by 61.5% and 74.6% on average compared to six state-of-the-art methods, respectively, and inference latency is reduced by a factor of 3.02. The solution has been successfully deployed and validated on NVIDIA Jetson platforms and real-world production lines.

Abstract

In modern high-throughput industrial production lines, product configurations and visual characteristics frequently change, making it impractical to collect and annotate data for every new scenario. This dynamic setting makes Zero-Shot Anomaly Detection (ZSAD) particularly suitable, as it enables defect detection without requiring training on target-specific samples. Although recent ZSAD approaches show promising results, they are computationally intensive and thus unsuitable for deployment on resource-constrained devices. We propose LiZAD: a lightweight framework designed for real-time ZSAD specifically tailored for use on edge devices. The proposed approach pairs the dense and spatially aware visual features of DINOv3, crucial for precise pixel-level localization, with the highly computationally efficient text embeddings of MobileCLIP2. These features are then mapped into a shared latent space via low-memory trainable projection heads. Compared to six state-of-the-art ZSAD models, LiZAD achieves an average memory reduction of 61.5%, a parameter reduction of 74.6%, and a speedup of 3.02x in terms of latency. Despite substantial reductions in computational and memory costs, our approach maintains competitive anomaly detection performance, dropping the average P-AUROC by just 6.4% relative to the best state-of-the-art model across the VisA, BTAD, MPDD, and MVTec-AD datasets. Finally, it is successfully deployed on the NVIDIA Jetson NX and Jetson AGX edge devices and tested on the real production line of the Industrial Computer Engineering Laboratory (ICE Lab) at the University of Verona. The code is available at https://github.com/intelligolabs/LiZAD.

Problem

• Zero-Shot Anomaly Detection
• Industrial Manufacturing
• Edge Devices
• Resource-Constrained Deployment
• Dynamic Product Configurations

Innovation

• Zero-Shot Anomaly Detection
• Lightweight Framework
• Edge Deployment
• DINOv3
• MobileCLIP