Particle size distribution based on deep learning instance segmentation
Baraian, Andrei-Cristian (2021-03-18)
Baraian, Andrei-Cristian
A.-C. Baraian
18.03.2021
© 2021 Andrei-Cristian Baraian. Tämä Kohde on tekijänoikeuden ja/tai lähioikeuksien suojaama. Voit käyttää Kohdetta käyttöösi sovellettavan tekijänoikeutta ja lähioikeuksia koskevan lainsäädännön sallimilla tavoilla. Muunlaista käyttöä varten tarvitset oikeudenhaltijoiden luvan.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:oulu-202103211425
https://urn.fi/URN:NBN:fi:oulu-202103211425
Tiivistelmä
Deep learning has become one of the most important topics in Computer Science, and recently it proved to deliver outstanding performances in the field of Computer Vision, ranging from image classification and object detection to instance segmentation and panoptic segmentation. However, most of these results were obtained on large, publicly available datasets, that exhibit a low level of scene complexity. Less is known about applying deep neural networks to images acquired in industrial settings, where data is available in limited amounts. Moreover, comparing an image-based measurement boosted by deep learning to an established reference method can pave the way towards a shift in industrial measurements.
This thesis hypothesizes that the particle size distribution can be estimated by employing a deep neural network to segment the particles of interest. The analysis was performed on two deep neural networks, comparing the results of the instance segmentation and the resulted size distributions. First, the data was manually labelled by selecting apatite and phlogopite particles, formulating the problem as a two-class instance segmentation task. Next, models were trained based on the two architectures and then used for predicting instances of particles on previously unseen images. Ultimately, accumulating the sizes of the predicted particles would result in a particle size distribution for a given dataset.
The final results validated the hypothesis to some extent and showed that tackling difficult and complex challenges in the industry by leveraging state-of-the-art deep learning neural networks leads to promising results. The system was able to correctly identify most of the particles, even in challenging situations. The resulted particle size distribution was also compared to a reference measurement obtained by the laser diffraction method, but still further research and experiments are required in order to properly compare the two methods. The two evaluated architectures yielded great results, with relatively small amounts of annotated data.
This thesis hypothesizes that the particle size distribution can be estimated by employing a deep neural network to segment the particles of interest. The analysis was performed on two deep neural networks, comparing the results of the instance segmentation and the resulted size distributions. First, the data was manually labelled by selecting apatite and phlogopite particles, formulating the problem as a two-class instance segmentation task. Next, models were trained based on the two architectures and then used for predicting instances of particles on previously unseen images. Ultimately, accumulating the sizes of the predicted particles would result in a particle size distribution for a given dataset.
The final results validated the hypothesis to some extent and showed that tackling difficult and complex challenges in the industry by leveraging state-of-the-art deep learning neural networks leads to promising results. The system was able to correctly identify most of the particles, even in challenging situations. The resulted particle size distribution was also compared to a reference measurement obtained by the laser diffraction method, but still further research and experiments are required in order to properly compare the two methods. The two evaluated architectures yielded great results, with relatively small amounts of annotated data.
Kokoelmat
- Avoin saatavuus [31941]