Training Based Segmentation of Electron Microscope Images
Valkonen, Masi Filemon (2015)
2015
Signaalinkäsittelyn ja tietoliikennetekniikan koulutusohjelma
Tieto- ja sähkötekniikan tiedekunta - Faculty of Computing and Electrical Engineering
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2015-09-09
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201508111509
https://urn.fi/URN:NBN:fi:tty-201508111509
Tiivistelmä
A great deal of image analysis in today’s materials research is done manually, which can be time consuming and tedious. This thesis is a case study of two image analysis problems in the field of materials science for which automatic methods are developed to aid the analysis process.
In the first case, an automatic segmentation method is developed to segment zeolite pores. The method is based on logistic regression combined with sparsity promoting LASSO regularization. It is able to find more pores than humans with better consistency and speed yielding average PAS score of 0.79 and F1 score of 0.89. Much of the error is caused by the fact that humans specify the pore perimeters in various ways, but the method consistently produces similar segmentations.
The second case considers automatic segmentation of silver nanoparticles by combining LASSO regularized logistic regression and watershed segmentation. The method is faster than manual segmentation and produces fine segments when particles have little overlap. However, if the amount of overlap is high, the segments are flawed. The performance of the system in terms of PAS metric and F1 score is 0.76 and 0.86, respectively.
An interesting property of training based segmentation together with sparsity promoting property is that training data can be collected by anyone. This enables creating an adaptive segmentation software for anyone regardless of image processing experience.
In the first case, an automatic segmentation method is developed to segment zeolite pores. The method is based on logistic regression combined with sparsity promoting LASSO regularization. It is able to find more pores than humans with better consistency and speed yielding average PAS score of 0.79 and F1 score of 0.89. Much of the error is caused by the fact that humans specify the pore perimeters in various ways, but the method consistently produces similar segmentations.
The second case considers automatic segmentation of silver nanoparticles by combining LASSO regularized logistic regression and watershed segmentation. The method is faster than manual segmentation and produces fine segments when particles have little overlap. However, if the amount of overlap is high, the segments are flawed. The performance of the system in terms of PAS metric and F1 score is 0.76 and 0.86, respectively.
An interesting property of training based segmentation together with sparsity promoting property is that training data can be collected by anyone. This enables creating an adaptive segmentation software for anyone regardless of image processing experience.