Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Multi-Dimensional data structures and classification schemes as a tool for the nondestructive analysis of complex materials

Kargakos Stefanos

Full record

Year 2022
Type of Item Diploma Work
Bibliographic Citation Stefanos Kargakos, "Multi-Dimensional data structures and classification schemes as a tool for the nondestructive analysis of complex materials", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2022
Appears in Collections
Relations with other Items


Microscopy is a critical component for the visualization of samples & objects thatcannot be seen with the unaided eye, allowing scientists to have a glimpse at a world ofunimaginable complexity. Different specialized microscopes deploying several imagingtechniques are essential for this purpose, however this way scientists tend to consume an unrealistic amount of time to achieve accurate diagnosis. In this thesis we exploit a new advanced HTS microscope featuring Hyperspectral, Transmission, Reflectance, Fluores-cence and Polarization imaging. By illustrating the morphological, molecular, electronic and crystalline structures of matter, this microscope provides unique features for auto-mated state of the art object classification. Upon observing those unique characteristics we approach a Multi-Modal object classification method, utilizing Convolutional Neural Networks for each modality and a Fully-Connected Neural Network which combines every unique illustration of the specimens. The CNN’s outputs are providing a serial unique encoding for each image and the FC-NN serve as a decoder capable of processing tabular labeled data. This thesis provides extensive analysis and results regarding the different combinations of imaging modalities with the intention to extract valuable information about their importance on the classification process. Exploiting the full power of our NNs system, by deploying every imaging modality we achieve accuracy greater than 99%.

Available Files