Artificial vision by thermography : calving prediction and defect detection in carbon fiber reinforced polymer

Authors: Fleuret, Julien
Advisor: Maldague, X.
Abstract: Abstract Computer vision is a field which consists in extracting or identifying one or more information from one or more images in order either to automate a task or to provide decision support. With the increase in the computing capacity of computers, the popularization and diversification of imaging means, both in industry, as well as in everyone’s life, this field has undergone a rapid development in recent decades. Among the different imaging modalities for which it is possible to use artificial vision, this thesis focuses on infrared imaging. More particularly on infrared imagery for wavelengths included in the medium and long bands. This thesis focuses on two radically different industrial applications. In the first part of this thesis, we present an application of artificial vision for the detection of the calving moment in industrial environments for Holstein cows. More precisely, the objective of this research is to determine the time of calving using only physiological data from the animal. To this end, we continuously acquired data on different animals over several days. Among the many challenges presented by this application, one of them concerns data acquisition. Indeed, the cameras we used are based on bolometric sensors, which are sensitive to a large number of variables. These variables can be classified into four categories: intrinsic, environmental, radiometric and geometric. Another important challenge in this research concerns the processing of data. Besides the fact that the acquired data uses a higher dynamic range than the natural images which complicates the processing of the data; Identifying recurring patterns in images and automatically recognizing them through machine learning is a major challenge. We have proposed a solution to this problem. In the rest of this thesis we have focused on the problem of defect detection in materials, using the technique of pulsed thermography. Pulse thermography is a very popular method due toits simplicity, the possibility of being used with a large number of materials, as well as its low cost. However, this method is known to produce noisy data. The main cause of this reputation comes from the various sources of distortion to which thermal cameras are sensitive. In this thesis, we have chosen to explore two axes. The first concerns the improvement of existing data processing methods. In the second axis, we propose several methods to improve fault detection. Each method is compared to several methods constituting the state of the art in the field.
Document Type: Thèse de doctorat
Issue Date: 2021
Open Access Date: 15 November 2021
Permalink: http://hdl.handle.net/20.500.11794/70927
Grantor: Université Laval
Collection:Thèses et mémoires

Files in this item:
Description SizeFormat 
37554.pdf32.64 MBAdobe PDFThumbnail
View/Open
All documents in CorpusUL are protected by Copyright Act of Canada.