Advancing unsupervised anomaly detection with normalizing flow and multi-scale ensemble learning
Visual anomaly detection plays a crucial role in manufacturing to ensure product quality by identifying image patterns that deviate from the expected ones. Existing methods that rely on distribution estimation struggle with the complexity of real-world images, resulting in complex and inefficient pr...
| Autores: | , , |
|---|---|
| Tipo de recurso: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universidad de Sevilla (US) |
| Repositorio: | idUS. Depósito de Investigación de la Universidad de Sevilla |
| OAI Identifier: | oai:idus.us.es:11441/165589 |
| Acceso en línea: | https://hdl.handle.net/11441/165589 https://doi.org/10.1016/j.engappai.2024.109088 |
| Access Level: | acceso abierto |
| Palabra clave: | Anomaly detection Novelty detection Deep learning Normalizing flow Unsupervised learning |
| Sumario: | Visual anomaly detection plays a crucial role in manufacturing to ensure product quality by identifying image patterns that deviate from the expected ones. Existing methods that rely on distribution estimation struggle with the complexity of real-world images, resulting in complex and inefficient procedures. This study leverages normalizing flow techniques to address the cold start anomaly detection problem, where no prior examples of anomalies are available during the training phase. In such scenarios, models must learn exclusively from defect-free images and still accurately identify anomalies. We propose a novel unsupervised multi-scale and multi-semantic normalizing flow model, enhanced with an ensemble of neural networks, to detect anomalies based on their feature distributions. Our model estimates the likelihood of non-defective features, identifying anomalies as out-of-distribution values. Extensive experiments on three state-of-the-art anomaly detection datasets demonstrate that our proposal achieves superior AUROC performance and improves computational efficiency compared to existing approaches. Furthermore, we validate the robustness and adaptability of our proposal through low-shot training experiments using only 20% of available training data, highlighting its potential as an efficient solution for cold start anomaly detection. |
|---|