guojiajeremy/Dinomaly

What it solves

Dinomaly addresses the performance gap in multi-class unsupervised anomaly detection (UAD). Traditionally, models trained to detect anomalies across multiple different classes of objects simultaneously performed significantly worse than models trained for a single specific class. Dinomaly provides a unified framework that can compete with, and sometimes exceed, the performance of single-class state-of-the-art models.

How it works

Dinomaly uses a minimalistic reconstruction-based approach built on pure Transformer architectures (Attentions and MLPs) without complex additional modules. It relies on four key components:

1. Foundation Transformers: Used to extract universal and discriminative features.
2. Noisy Bottleneck: Utilizes existing Dropouts to inject noise into the process.
3. Linear Attention: A mechanism that prevents the model from focusing too narrowly.
4. Loose Reconstruction: A strategy that avoids forcing strict point-by-point or layer-to-layer reconstruction, allowing for more flexible anomaly detection.

Who it’s for

This project is designed for researchers and engineers working on computer vision, specifically those implementing industrial quality control or anomaly detection systems where a single model must handle multiple different product categories.

Highlights

• High Performance: Achieves image-level AUROC of up to 99.8% on MVTec-AD, 98.9% on VisA, and 90.1% on Real-IAD.
• Scalable: Performance scales effectively with model size, input image resolution, and the choice of foundation model.
• Minimalist Design: Avoids specialized tricks or complex modules in favor of a clean Transformer-based architecture.
• Integrated: Now integrated into the Intel open-edge Anomalib library (v2.1.0).