Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Development of machine learning techniques towards spectral dimensionalityexpansion : applications in snapshot spectral imaging.

Logothetis Fragkoulis

Full record

Year 2020
Type of Item Master Thesis
Bibliographic Citation Fragkoulis Logothetis, "Development of machine learning techniques towards spectral dimensionality expansion : applications in snapshot spectral imaging.", Master Thesis, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2020
Appears in Collections


Hyperspectral imaging (HSI) is an emerging technology that integrates conventionalimaging and spectroscopy to attain both spatial and spectral informationfrom an object. The spectral images, collected in the spectral cube, are tensof hundreds and the information we receive from them is crucial for many applications,such as bio-medical technology, remote sensing, microscopy etch.Nevertheless, the current state of the art includes HSI systems which needlong acquisition time, something that prevents them from observing any dynamicallydeveloping phenomena. Also, they are expensive and sizeable, whichmakes them inaccessible to many important applications. To address theselimitations, a theoretical real-time snapshot spectral imaging system (SNSI)that captures a small number of spectral bands, and by using dimensionalityexpansion techniques, provides real time HSI, is investigated in this study. Acomparative study of the state of the art was conducted under the assumptionof various hardware architectures (RGB narrow/wide, three, six, nine, andtwelve spectral channels). Furthermore, two new algorithms called K-Fourierand 2Level are proposed and compared in terms of minimizing the estimationerror of the uncaptured spectral images. The novelty of the proposed methodsstems mainly from the reduction of the dimensionality of the space needed to bereconstructed. Experiments on standard color charts show that the K-Fourierand non-linear kernels outperform the other competing methods. Looking at thesame problem from another perspective, the most feature-rich training yields tohigher estimation accuracy. On that account, a great amount of band selectiontechniques, which are based on similarity-measurement, dynamic programming,and evolutionary formulas were analyzed and compared. Genetic Algorithmsturned out to be the most promising feature selection technique, since it dramaticallyimproves the space reconstruction error. Moreover, we introduce anonlinear transformation of reflectance values to ensure that the estimated reflectionspectra fulfill physically motivated boundary conditions. Ending up,these findings set the basis for the development of a powerful SNSI system.Medical diagnosis is expected to be a leading application of this novel approach.

Available Files