This thesis addresses the problem of detecting novel and unforeseen situations in computer vision systems based on deep learning models, with particular focus on the phenomena of concept drift and novelty detection. In many real-world application scenarios, models are trained on a limited dataset and subsequently deployed in dynamic environments, where objects, configurations, or operating conditions may arise that differ from those observed during the training phase, thereby compromising the reliability of predictions. The main objective of this work is the design and implementation of a Concept Drift Detection module capable of automatically recognizing such changes and signaling the presence of potentially anomalous or semantically novel observations. To this end, pre-trained convolutional neural networks are employed as feature extractors, leveraging compact and discriminative latent representations of the input images. Novelty detection is performed in the feature space by comparing new observations against a reference distribution learned during an initial baseline phase. Techniques based on statistical distance measures and feature distribution analysis are explored, alongside clustering-based methods for outlier identification. The system outputs an anomaly score and an alert mechanism, providing useful support for model update decisions or further in-depth analysis. Experimental results demonstrate that the proposed approach effectively identifies semantic variations in the data, representing a robust and flexible first level of monitoring for computer vision and machine learning applications in real-world, non-stationary environments.
Questo lavoro di tesi affronta il problema del rilevamento di situazioni nuove e non previste in sistemi di computer vision basati su modelli di deep learning, con particolare attenzione ai fenomeni di \textit{concept drift} e \textit{novelty detection}. In molti scenari applicativi reali, i modelli vengono addestrati su un insieme limitato di dati e successivamente impiegati in ambienti dinamici, nei quali possono comparire oggetti, configurazioni o condizioni operative differenti rispetto a quelle osservate in fase di training, compromettendo l’affidabilità delle predizioni. L’obiettivo principale del lavoro è la progettazione e l’implementazione di un modulo di \textit{Concept Drift Detection} in grado di riconoscere automaticamente tali cambiamenti, segnalando la presenza di osservazioni potenzialmente anomale o semanticamente nuove. A tal fine, vengono utilizzate reti neurali convoluzionali pre-addestrate come estrattori di feature, sfruttando rappresentazioni latenti compatte e discriminative delle immagini di input. La \textit{novelty detection} viene effettuata operando nello spazio delle feature, confrontando le nuove osservazioni con una distribuzione di riferimento appresa durante una fase iniziale di baseline. Vengono esplorate tecniche basate su misure di distanza statistica e analisi della distribuzione delle feature, affiancate da metodi di clustering per l’individuazione di outlier. Il sistema restituisce uno score di anomalia e un meccanismo di alert, utile per supportare decisioni di aggiornamento del modello o analisi più approfondite. I risultati sperimentali mostrano come l’approccio proposto consenta di individuare efficacemente variazioni semantiche nei dati, rappresentando un primo livello di controllo robusto e flessibile per applicazioni di computer vision e machine learning in contesti reali e non stazionari.
PROGETTAZIONE E SVILUPPO DI UN SISTEMA DI CONCEPT DRIFT DETECTION PER SCENARI DI VISIONE ARTIFICIALE
MARRONE, ANTONELLA
2024/2025
Abstract
This thesis addresses the problem of detecting novel and unforeseen situations in computer vision systems based on deep learning models, with particular focus on the phenomena of concept drift and novelty detection. In many real-world application scenarios, models are trained on a limited dataset and subsequently deployed in dynamic environments, where objects, configurations, or operating conditions may arise that differ from those observed during the training phase, thereby compromising the reliability of predictions. The main objective of this work is the design and implementation of a Concept Drift Detection module capable of automatically recognizing such changes and signaling the presence of potentially anomalous or semantically novel observations. To this end, pre-trained convolutional neural networks are employed as feature extractors, leveraging compact and discriminative latent representations of the input images. Novelty detection is performed in the feature space by comparing new observations against a reference distribution learned during an initial baseline phase. Techniques based on statistical distance measures and feature distribution analysis are explored, alongside clustering-based methods for outlier identification. The system outputs an anomaly score and an alert mechanism, providing useful support for model update decisions or further in-depth analysis. Experimental results demonstrate that the proposed approach effectively identifies semantic variations in the data, representing a robust and flexible first level of monitoring for computer vision and machine learning applications in real-world, non-stationary environments.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/109505