TY - THES ID - 146384098 TI - Master thesis : Machine learning and Multispectral Imaging for Art conservation AU - Tasset, Maxime AU - Marée, Raphaël AU - Geurts, Pierre AU - Wehenkel, Louis AU - Van Droogenbroeck, Marc AU - Lasaracina, Karine PY - 2018 PB - Liège Université de Liège (ULiège) DB - UniCat KW - painting KW - cytomine KW - multi spectral KW - image KW - crack KW - Ingénierie, informatique & technologie > Sciences informatiques UR - https://www.unicat.be/uniCat?func=search&query=sysid:146384098 AB - Detecting cracks is a very interesting field of research used to address a large number of issues, including the detection of road and other structural cracks to strengthen the prevention of damages and planning of maintenance and repairs. It is also used to facilitate the authentication of paintings. In most cases, the detection is only performed on simple images while multi-spectral images are increasingly used in many research areas, including Medicine, given their high potential for providing more accurate information. This work responds to a request by the "Lumiere Technology Multispectral Institute" which indicated an interest in using multi-spectral images to detect cracks in paintings. Its objective is to analyse if applying detections on multi-spectral images can help improve the quality and efficiency of maintenance and restoration of paintings. To meet this objective, we work with the Cytomine application, an "Open-source rich Internet application for collaborative analysis of multi-gigapixel images", developed by researchers at the Montefiore Institute of the University of Liège. We more specifically use the recently-developed extension supporting the utilisation of multi-spectral images to apply machine learning algorithms on wide-banded multi-spectral images. The "Musées royaux des beaux-arts de Belgique" have provided the paintings which are analysed in this work. To carry out our analysis: i) we implement the Cytomine’s Extractor and Spectral Reader and complement the application’s new extension by adding ways to optimise the extraction, storage and use of data from multi-spectral image, ii) based on this implementation, we present three softwares which we designed to enable us to make a feature analysis of the data, to fit a model using data from extit{Cytomine}, or to use a fitted model to make predictions, iii) we test two dimension reduction methods, PCA (Principal Components Analysis) and TSNE (T-distributed Stochastic Neighbour Embedding) and three feature importance measures to help reducing the large size of datasets created by multi-spectral images, and iv) we conduct experiments on three datasets to find how machine learning performs on multi-spectral images: one small dataset based on a multi-spectral image of biological cells, one small dataset based on a multi-spectral image where we have extracted two different tones of red, and a third dataset based on a multi-spectral image where we have labelled cracks and undamaged parts on the painting extit{Portrait of Flautist François Devienne} of Jacques-Louis David. These experiments evidence that: i) TSNE is a possible solution to reduce the dimension of multi-spectral images if the necessary resources, i.e. memory and computational power, are available, and ii) using multi-spectral images has an advantage over simply using standard RGB images, even if the increase in computation time implies that a pre-processing is needed to reduce the number of bands. We therefore encourage further research based on the new tools developed in this work, for instance to try Boosting methods or Deep Neural Networks. ER -