Year: 2021
Type: dissertation
Abstract: Transformers have been getting very popular in Natural Language Processing since their introduction in "Attention Is All You Need" (2017), more recently, after the introduction of the Vision Transformer [Dosovitskiy et al. (2020)] they have started to gain popularity also in the field of computer vision for what concerns image classification. In this work, it is proposed a visual anomaly detection system based on attention. For this kind of task, the Convolutional Neural Networks (CNNs) represent the dominant approach and they are the state-of-the-art. The proposed architecture, based on Vision Transformers and attention, achieved good results on the UCSD PED2 Anomaly Detection Dataset, even if it failed to outperform the state-of-the-art, the outcome is encouraging considering how new these technologies are. CNNs have been using for more than a decade and they are a fully mature technology, whereas Vision Transformers have been using only for several months, there are reasons to believe that in the future these approaches will be mature enough to become state-of-the-art.
Author: SCOTTO, FILIPPO
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7055841439