Agnostic object counting using deep learning for manufacturing applications

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en science des données, à finalité spécialisée: analyse de données

Abstract

enThe objective of this master's thesis is to explore deep learning techniques for agnostic object counting in manufacturing environments. The study begins by examining current object counting methods and gradually introduces class-agnostic counting. A comprehensive comparison methodology is presented to evaluate these methods, using specific datasets and evaluation metrics to meet manufacturing requirements. A significant part of the thesis is dedicated to developing a customized model that aims to strike a balance between performance and efficiency. This work denotes three primary contributions: the adaptation of existing models to better suit industrial applications, the introduction of a new dataset (FSCindu) comprising images closer to manufacturing scenarios, and the integration of different techniques to improve model performance. The research also addresses challenges such as handling objects with varying aspect ratios and optimizing the model for real-time processing. The results demonstrate significant improvements in object counting accuracy and efficiency, competing effectively with current models in this trade-off. However, some limitations are also present, of which the manufacturing sector must be aware. The thesis concludes with a discussion on future directions and potential advancements in the field. Code is available here [https://github.com/Diaby02/EasyCounting].