TY - THES

T1 - Digital Image Analysis of oxide particle dissolution in slags observed by HT-LSCM

AU - Lenzhofer, Florian

N1 - embargoed until 21-06-2027

PY - 2022

Y1 - 2022

N2 - Dissolution experiments of refractory particles in different slags using HT-LSCM are conducted aiming to determine effective diffusion coefficients of the refractory components in these slags. An evaluation of the recorded videos of these dissolution processes is needed to gain the particle diameter decrease with dissolution time. This evaluation is currently done manually. The biggest drawback of this method is that only a few frames are evaluated. To gain a higher number of evaluated frames, a at least partly automated procedure using digital image analysis should be established. At first, an analysis of the data sets was performed. A literature research was conducted to find suitable image enhancement and segmentation techniques. Due to permanent change of parameters most of the common methods seemed to be unusable. The best results were given by the Felzenszwalb-algorithm, an efficient graph-based image segmentation method. Since the image properties are changing during the experiment and between the different experiments, this algorithm soon reaches its limits. The next step was to use the open-source software ilastik for image classification and segmentation. For the data received from the dissolution experiments, the implemented classifier of ilastik was unfortunately insufficient. But it becomes an excellent tool to manually label frames. These manually labeled frames were used as training data for the U-Net model, which was finally implemented for the image analysis. More than 5000 manually segmented frames were used for training the U-Net model. The model was applied to three different experiments to predict the particle diameter decrease with dissolution time. For one of the evaluated experiments prediction failed; here bubble formation was the main reason. For the second experiment a deviation of the prediction to manually evaluated particle diameter decrease was observed while for the last one a satisfying prediction was achieved. The main challenge for the segmentation of these data sets is the change of several parameters during the experiment. The objective magnification of the microscope is not constant. The partly different illumination is caused by the experimental setup. In some cases, there are bubbles that look like the particle. The particle itself changes its shape, size and visibility. Due to the particle movement the background is changing during the experiment. Due to the described complexity of the data, a satisfying segmentation result is not possible for all experiments. Possible solutions may be the handling with the hyperparameter of the model. A second possibility is to separate the data and train particular models for data sets with equal properties. Perhaps another deep learning model leads to better results.

AB - Dissolution experiments of refractory particles in different slags using HT-LSCM are conducted aiming to determine effective diffusion coefficients of the refractory components in these slags. An evaluation of the recorded videos of these dissolution processes is needed to gain the particle diameter decrease with dissolution time. This evaluation is currently done manually. The biggest drawback of this method is that only a few frames are evaluated. To gain a higher number of evaluated frames, a at least partly automated procedure using digital image analysis should be established. At first, an analysis of the data sets was performed. A literature research was conducted to find suitable image enhancement and segmentation techniques. Due to permanent change of parameters most of the common methods seemed to be unusable. The best results were given by the Felzenszwalb-algorithm, an efficient graph-based image segmentation method. Since the image properties are changing during the experiment and between the different experiments, this algorithm soon reaches its limits. The next step was to use the open-source software ilastik for image classification and segmentation. For the data received from the dissolution experiments, the implemented classifier of ilastik was unfortunately insufficient. But it becomes an excellent tool to manually label frames. These manually labeled frames were used as training data for the U-Net model, which was finally implemented for the image analysis. More than 5000 manually segmented frames were used for training the U-Net model. The model was applied to three different experiments to predict the particle diameter decrease with dissolution time. For one of the evaluated experiments prediction failed; here bubble formation was the main reason. For the second experiment a deviation of the prediction to manually evaluated particle diameter decrease was observed while for the last one a satisfying prediction was achieved. The main challenge for the segmentation of these data sets is the change of several parameters during the experiment. The objective magnification of the microscope is not constant. The partly different illumination is caused by the experimental setup. In some cases, there are bubbles that look like the particle. The particle itself changes its shape, size and visibility. Due to the particle movement the background is changing during the experiment. Due to the described complexity of the data, a satisfying segmentation result is not possible for all experiments. Possible solutions may be the handling with the hyperparameter of the model. A second possibility is to separate the data and train particular models for data sets with equal properties. Perhaps another deep learning model leads to better results.

KW - HT-LSCM

KW - U-NET

KW - digitale Bildanalyse

KW - Auflösung oxidischer Partikel

KW - HT-LSCM

KW - U-NET

KW - image analysis

KW - oxide particle dissolution

M3 - Master's Thesis

ER -