Instance Segmentation for Rock Particle Quality Monitoring: Integration of Deep Learning for Machine Vision Application in the Aggregates Industry
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2024-01-22
Department
Major/Subject
Computer Science
Mcode
SCI3042
Degree programme
Master’s Programme in Computer, Communication and Information Sciences
Language
en
Pages
63
Series
Abstract
Following the new directives of the European Green Deal, many industries, including the aggregates industry, are facing new challenges: the need for more sustainable, robust and cost-efficient solutions is growing. The analysis and classification of rocks travelling on conveyor belts between processes provides the means to optimize production. Therefore, it is reducing the environmental impact and costs, while increasing the efficiency of the processes. Metso attempts to offer an optimization of the rock particle quality monitoring systems by integrating deep learning for machine vision applications. The purpose is to provide a more accurate, low-cost and easily redeployable solution that could be used in the aggregates industry or elsewhere. Prior to this study, implementations were attempted using traditional machine vision or deep learning object detection methods. However, the results were unsatisfactory. The goal of this thesis is to improve the current solution by implementing instance segmentation into the rock particles monitoring system. It also separates the extraction of application-specific statistics from the general detection of particles in order to increase the redeployability of the product. Furthermore, it compares the results of different deep learning frameworks to explore potential approaches for future development. The evaluation of the proposed solution displayed acceptable results with greater accuracy and speed. The calculations of the rock statistics increased the redeployability of the system with clear separation from the detection models. The tests of the models from Annonet, YOLOv8 and Roboflow frameworks highlighted the advantages of the instance segmentation approach. Moreover, the YOLOv8 models outperformed the others in all the metrics, noticeably in terms of speed. In particular, the YOLOv8 models showed promising results on both CPU and GPU-based applications, thus, potentially further reducing the costs of the system.Description
Supervisor
Solin, ArnoThesis advisor
Palokangas, JaakkoKeywords
deep learning, neural networks, instance segmentation, object detection, machine vision, aggregates