Feedback

Faculté des Sciences appliquées
Faculté des Sciences appliquées
MASTER THESIS
VIEW 73 | DOWNLOAD 15

Annotating artistic images with deep learning

Download
Noirhomme, Maxime ULiège
Promotor(s) : Geurts, Pierre ULiège
Date of defense : 26-Jun-2019/27-Jun-2019 • Permalink : http://hdl.handle.net/2268.2/6774
Details
Title : Annotating artistic images with deep learning
Author : Noirhomme, Maxime ULiège
Date of defense  : 26-Jun-2019/27-Jun-2019
Advisor(s) : Geurts, Pierre ULiège
Committee's member(s) : Van Droogenbroeck, Marc ULiège
Marée, Raphaël ULiège
Language : English
Number of pages : 72
Discipline(s) : Engineering, computing & technology > Computer science
Institution(s) : Université de Liège, Liège, Belgique
Degree: Master en ingénieur civil en informatique, à finalité spécialisée en "intelligent systems"
Faculty: Master thesis of the Faculté des Sciences appliquées

Abstract

[en] In this work, we study the problem of finding bounding boxes around musical instruments in artistic images. We do this by training deep neural networks on datasets which represent different instruments within natural photographs, and test their performance within paintings that represent the same kind of instruments. Our goal is to first train a classifier that is able of classifying different representations of the musical instruments, and to then use this classifier as a feature extractor for finding bounding boxes.

For the training of the classifier, we use Domain Adversarial Neural Networks (DANN) and compare their performance to more classical transfer learning approaches. Furthermore we investigate the use of style transfer techniques as a way of data augmentation. This allows us to fill the gap between images represented in a photo-realistic way and the ones in paintings, and could overall help reducing the generalisation error of our trained models.

Our results show that when it comes to the artistic images of the instruments, DANNs appear to perform as poorly as classic transfer learning techniques, and that our way of performing data augmentation does not significantly reduce the generalisation error.

However, on a simplified version of the problem we show that good performance can be achieved at least for the classification part, while for the detection of the bounding boxes around the images more work is required.


File(s)

Document(s)

File
Access Master_thesis_final.pdf
Description: Master Thésis
Size: 20.56 MB
Format: Adobe PDF
File
Access Résumé.pdf
Description: -
Size: 47.7 kB
Format: Adobe PDF

Author

  • Noirhomme, Maxime ULiège Université de Liège > Master ingé. civ. info., à fin.

Promotor(s)

Committee's member(s)

  • Van Droogenbroeck, Marc ULiège Université de Liège - ULiège > Dép. d'électric., électron. et informat. (Inst.Montefiore) > Télécommunications
    ORBi View his publications on ORBi
  • Marée, Raphaël ULiège Université de Liège - ULiège > Dép. d'électric., électron. et informat. (Inst.Montefiore) > Systèmes et modélisation
    ORBi View his publications on ORBi
  • Total number of views 73
  • Total number of downloads 15










All documents available on MatheO are protected by copyright and subject to the usual rules for fair use.
The University of Liège does not guarantee the scientific quality of these students' works or the accuracy of all the information they contain.