Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Development of supervised and unsupervised methods for spectral estimation in hyper-spectral imaging

Logothetis Fragkoulis

Full record

---

Summary

Hyperspectral imaging (HSI) is an emerging technology that integrates conventional imaging and spectroscopy to attain both spatial and spectral information from an object. The spectral images collected in the spectral cube are tens of 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 need long acquisition time, something that prevents them from observing any dynamically developing phenomena. Also, are expensive and sizeable, which makes them inaccessible for many important applications. To address these limitations, a real-time snapshot spectral imaging system (SNI) that captures only six spectral bands, and by using dimensionality expansion techniques provides real time HIS was developed by us. To achieve that goal, a plenty of spectral estimation methods were leveraged and compared in terms of minimizing the estimation error of the uncaptured spectral images. Wiener, PCA, Random Forest Regression, Linear Regression, Neural Networks, and Support vector Regression are some of the methods that were applied to expand the dimensionality of the spectral domain. To do this more accurately, we propose a new Adaptive Hybrid method that overperforms the above spectral estimation models and achieves the lowest estimation error. The novelty of the proposed method stems mainly both from the combination of Wiener and PCA models and from the manner of the training sample selection. Looking at the same problem from another perspective, in order to increase the performance of the models, training procedure should be carried out using the most informative and distinctive bands as features. On that account, a great amount of band selection techniques, which are based either on geometric approaches or on similarity-based formulas were analyzed and compared. The main shortcoming is that there is no band extraction method specific for spectral estimation approaches. For that reason, a new band selection method specific for the spectral dimensionality expansion was devised by us. Experimental results show that the proposed method locate the most featured bands, which decrease dramatically the root mean square error between the reference and the estimated spectrum, compared with the well-known OSP method. These findings were set the basis for the development of a powerful SNI.

Available Files

Services