This thesis presents a novel method for anomaly detection and localization in image data, leveraging denoising diffusion probabilistic models. Traditional methods in industrial quality control often need help with precision and efficiency, which this research aims to address through an advanced diffusion-based approach. The proposed method involves a two-step process: a forward process that introduces noise to images and a reverse process that denoises them, effectively reconstructing the original images and identifying anomalies. The core innovation lies in using self-supervised learning to reduce dependency on labeled datasets, which are typically expensive and time-consuming. This method generates synthetic anomalies for training, enhancing model robustness and generalization. Experimental results demonstrate significant improvements in anomaly detection accuracy over state-of-the-art methods validated on the benchmark MVTec AD dataset. The thesis also explores architectures like U-Net to optimize the anomaly segmentation process, comprehensively evaluating their performance. This work contributes to the field by offering a scalable, efficient, and highly accurate solution for real-time anomaly detection in industrial settings.
Questa tesi presenta un metodo innovativo per il rilevamento e la localizzazione di anomalie nei dati di immagine, sfruttando modelli probabilistici di diffusione per la riduzione del rumore. I metodi tradizionali nel controllo di qualità industriale spesso incontrano difficoltà in termini di precisione ed efficienza, problemi che questa ricerca intende affrontare attraverso un approccio avanzato basato sulla diffusione. Il metodo proposto prevede un processo in due fasi: un processo forward che introduce rumore nelle immagini e un processo reverse che riduce il rumore, ricostruendo efficacemente le immagini originali e identificando le anomalie. L'innovazione principale risiede nell'utilizzo dell'apprendimento auto-supervisionato per ridurre la dipendenza dai dataset etichettati, che sono tipicamente costosi e richiedono molto tempo. Questo metodo genera anomalie sintetiche per l'addestramento, migliorando la robustezza e la generalizzazione del modello. I risultati sperimentali dimostrano miglioramenti significativi nell'accuratezza del rilevamento delle anomalie rispetto ai metodi all'avanguardia, validati sul dataset di riferimento MVTec AD. La tesi esplora anche architetture come U-Net per ottimizzare il processo di segmentazione delle anomalie, valutandone in modo completo le prestazioni. Questo lavoro contribuisce al campo offrendo una soluzione scalabile, efficiente e altamente accurata per il rilevamento delle anomalie in tempo reale negli ambienti industriali.
Anomaly Detection in Image Data using Denoising Diffusion Probabilistic Models
AZAD, FATEMEH
2023/2024
Abstract
This thesis presents a novel method for anomaly detection and localization in image data, leveraging denoising diffusion probabilistic models. Traditional methods in industrial quality control often need help with precision and efficiency, which this research aims to address through an advanced diffusion-based approach. The proposed method involves a two-step process: a forward process that introduces noise to images and a reverse process that denoises them, effectively reconstructing the original images and identifying anomalies. The core innovation lies in using self-supervised learning to reduce dependency on labeled datasets, which are typically expensive and time-consuming. This method generates synthetic anomalies for training, enhancing model robustness and generalization. Experimental results demonstrate significant improvements in anomaly detection accuracy over state-of-the-art methods validated on the benchmark MVTec AD dataset. The thesis also explores architectures like U-Net to optimize the anomaly segmentation process, comprehensively evaluating their performance. This work contributes to the field by offering a scalable, efficient, and highly accurate solution for real-time anomaly detection in industrial settings.File | Dimensione | Formato | |
---|---|---|---|
Azad_Fatemeh.pdf
embargo fino al 08/01/2026
Dimensione
31.58 MB
Formato
Adobe PDF
|
31.58 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/66526