Work

Aprenentatge Semi-Supervisat per la Segmentació d'Imatges

Year: 2021
Type: dissertation
Abstract: In the 21st century, technological challenges are becoming more and more ambitious and diverse. In this thesis, we will use one relatively new technique whose whole potential is yet to be discovered, Machine Learning (ML), and its great predictive capacity. ML has a wide spectrum of features and applications, and to reach relevant results, this project focuses on the narrow but ambitious objective of segmenting images, a powerful concept that has a direct application to diverse sectors, such as autonomous driving or medical imaging. Thanks to the fast development and extension in the society of multimedia technologies, the amount of data we have to work with this type of algorithm is immense. However, the process of labelling these images (or videos) is highly costly, or even in some cases, is not feasible. Recently, the scientific community has shown a possibility that is giving extraordinary results, known as semi-supervised learning. This technique consists of using a few data to make a supervised training of the system, and later, to use unsupervised techniques to improve significantly the predictions. In this project, firstly, the FixMatch algorithm is reimplemented for the image classification task, achieving an accuracy of 90.11%, secondly, many image segmentation models have been analyzed and tested, and lastly, FixMatch has been adapted to the task of image segmentation giving a first approach to the coming researches in the field of semi-supervised semantic segmentation.
Author: López Serrano, Jordi
Language: English
Open Access status: green
OpenAlex ID: W7113811119
Cites: 59
Cited by:
Related to: 10