8 Ekim 2012 Pazartesi

Efficient Segmentation and Classification of Hyper Spectral Cubes

Efficient Segmentation and Classification of Hyper Spectral Cubes

Papagrigoriou Stylianos
Technical University of Crete


Hyper Spectral Imaging is a powerful analytical tool, which has been used in a wide area of applications, from Satellite Imaging to Biomedical Diagnosis. Spectral imagery of either macroscopic or microscopic origin is usually depicted in a spectral cube, a registered set of images, featuring one spectral and two spatial dimensions as pixel coordinates. From each pixel, associated with a spectrum -instead of an RGB value-, one is able to extract information about the nature of the material, by studying its spectral signature on the Spectral Cube.
This technique offers a non-destructive and non-invasive way (one does not have to extract part of the material and bring it to the lab) of examining materials, suitable for medical purposes. In the hereby thesis the computational capabilities of spectral imaging methods are examined and attempted to be improved, in order to provide real time pixel classification. Specifically, a successful attempt is made to create a hyper spectral classifier with real-time performance for cubes acquired from a cervix biopsy. Various techniques are tested for efficient segmentation of the Cube, in order to generate a golden standard for the training process. The classification is performed using Neural Networks while the final result is a GPU implementation, the main reason behind the speed up of the application.
Although this study was based on specific medical data, it is possible to be generalized on any aspect of Hyper Spectral Imaging, and shows that real-time Hyper Spectral Processing for classification purposes is feasible.

FannTool ;

FANN library, provides a really fast and powerful tool -FannTool- for training neural networks with many different algorithms and automatic mechanisms for identifying the best activation function and training algorithm. FANN library was used exactly for these two purposes (finding the best activation function and training algorithm) and train the Network

0 yorum:

Yorum Gönder