Semantic segmentation is a fundamental task in computer vision, essential for various applications requiring precise object delineation. A significant challenge in this domain is the substantial labeling effort required to train accurate models. This thesis explores methods to reduce the labeling burden by implementing active learning (AL) and semi-supervised learning (SSL) techniques. In this study, an AL strategy was employed to identify the most informative samples for labeling. The effectiveness of this method was evaluated by comparing the accuracies obtained from labeling the AL-suggested samples with those achieved by selecting pre-labeled images based on their known accuracy after inference. This approach allowed the selection of images that were poorly inferred by the model, which are presumed to have a greater impact on improving model performance when added to the training set. The AL approach demonstrated excellent performance, yelding results very close to the baseline model trained on the “best” selection of the dataset. Additionally, one SSL strategy was explored to leverage the vast amounts of unlabeled data. Differently from the original paper, this work explores the application of the methodology in multi-labeling tasks. Moreover, two modified strategies were investigated: the first aimed at reducing computational effort, and the second focused on enhancing accuracy performance. Although these strategies are still being refined, initial findings suggest that they hold promise for reducing the labeling effort without compromising segmentation quality.
La segmentazione semantica è una tecnica fondamentale in computer vision, essenziale per una molteplicità di applicazioni che richiedono una delineazione precisa degli oggetti. In questo campo è gravoso il lavoro di etichettatura necessario per l’addestramento di modelli accurati. Questa tesi esplora metodi per ridurre questo sforzo implementando tecniche di apprendimento attivo (AL) e di apprendimento semi-supervisionato (SSL). In questo studio, è stata impiegata una strategia di AL per identificare i campioni più informativi da etichettare. L’efficacia di questo metodo è stata valutata confrontando le precisioni conseguite dopo aver etichettato i campioni suggeriti dall’algoritmo di AL con quelle ottenute mediante una selezione di immagini pre-etichettate, nota la loro precisione dopo l’inferenza. Questo approccio ha permesso di selezionare le immagini imprecisamente interpretate dal modello, assumendo abbiano un impatto maggiore sul miglioramento delle prestazioni del modello se aggiunte al set di addestramento. L’approccio di AL ha raggiunto prestazioni eccellenti, con risultati molto vicini al modello addestrato sulla “migliore” selezione del set di dati. Inoltre, per sfruttare la grande quantità di dati non etichettati, è stata esplorata una strategia di SSL. A differenza dell’approccio proposto nell’articolo originale, questa tesi esplora l’applicazione della metodologia in ambito multi-labeling. In aggiunta, sono state studiate due strategie modificate: la prima mirata a ridurre lo sforzo computazionale, la seconda volta a migliorare le prestazioni di accuratezza. Sebbene queste strategie siano ancora in fase di perfezionamento, i primi risultati suggeriscono che sono promettenti ed utili a ridurre lo sforzo di etichettatura senza compromettere la qualità della segmentazione.
Active and Semi-Supervised Learning for Semantic Segmentation in Parapharmaceutical Inspection: A Study on Reducing Labeling Workload.
ZANIN, DARIA
2023/2024
Abstract
Semantic segmentation is a fundamental task in computer vision, essential for various applications requiring precise object delineation. A significant challenge in this domain is the substantial labeling effort required to train accurate models. This thesis explores methods to reduce the labeling burden by implementing active learning (AL) and semi-supervised learning (SSL) techniques. In this study, an AL strategy was employed to identify the most informative samples for labeling. The effectiveness of this method was evaluated by comparing the accuracies obtained from labeling the AL-suggested samples with those achieved by selecting pre-labeled images based on their known accuracy after inference. This approach allowed the selection of images that were poorly inferred by the model, which are presumed to have a greater impact on improving model performance when added to the training set. The AL approach demonstrated excellent performance, yelding results very close to the baseline model trained on the “best” selection of the dataset. Additionally, one SSL strategy was explored to leverage the vast amounts of unlabeled data. Differently from the original paper, this work explores the application of the methodology in multi-labeling tasks. Moreover, two modified strategies were investigated: the first aimed at reducing computational effort, and the second focused on enhancing accuracy performance. Although these strategies are still being refined, initial findings suggest that they hold promise for reducing the labeling effort without compromising segmentation quality.File | Dimensione | Formato | |
---|---|---|---|
Zanin_Daria.pdf
accesso riservato
Dimensione
5.45 MB
Formato
Adobe PDF
|
5.45 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/20.500.12608/74385