Publication: Aceleración de algoritmos para imágenes hiperespectrales con OpenCL
Loading...
Official URL
Full text at PDC
Publication Date
2016
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
En el presente trabajo se propone dar solución a uno de los problemas principales surgido en el campo del análisis de imágenes hiperespectrales. En las últimas décadas este campo está siendo muy activo, por lo que es de vital importancia tratar su problema principal: mezcla espectral. Muchos algoritmos han tratado de solucionar este problema, pero que a través de este trabajo se propone una cadena nueva de desmezclado en paralelo, para ser acelerados bajo el paradigma de programación paralela de OpenCl. Este paradigma nos aporta el modelo de programación unificada para acelerar algoritmos en sistemas heterogéneos. Podemos dividir el proceso de desmezclado espectral en tres etapas. La primera tiene la tarea de encontrar el número de píxeles puros, llamaremos endmembers a los píxeles formados por una única firma espectral, utilizaremos el algoritmo conocido como Geometry-based Estimation of number of endmembers, GENE. La segunda etapa se encarga de identificar los píxel endmembers y extraerlos junto con todas sus bandas espectrales, para esta etapa se utilizará el algoritmo conocido por Simplex Growing Algorithm, SGA. En la última etapa se crean los mapas de abundancia para cada uno de los endmembers encontrados, de esta etapa será encargado el algoritmo conocido por, Sum-to-one Constrained Linear Spectral Unmixing, SCLSU. Las plataformas utilizadas en este proyecto han sido tres: CPU, Intel Xeon E5-2695 v3, GPU, NVidia GeForce GTX 980, Acelerador, Intel Xeon Phi 31S1P. La idea de este proyecto se basa en realizar un análisis exhaustivo de los resultados obtenidos en las diferentes plataformas, con el fin de evaluar cuál se ajusta mejor a nuestras necesidades.
In this work it intends give a solution to a problem hiperespectral images analysis field. This field is growing due to the analysis of hyperspectral images. Many algorithms have tried to solve this problem, this paper bring new algorithms to be accelerated under the OpenCl parallel programming paradigm. This paradigm give us the unified programming model to accelerate algorithms in heterogeneous systems. We can split the spectral unmixing in three stages. The first has the task of finding the number of pure pixels, called endmembers, pixels formed of a single spectral signature. Will be used the algorithm known as Geometry-based Estimation of number of endmember, GENE. The second stage is responsible for identifying endmember pixel and extract them along with all its bands. For this purpose we choose the algorithm called Simplex Growing Algorithm, SGA. In the last stage, making abundance maps of each endmember will be created. This step will be charged the algorithm known as, Sum-to-one Constrained Linear Spectral Unmixing, SCLSU. Three platforms have been used in this paper: As CPU, Intel Xeon E5-2695 v3, GPU, NVidia GeForce GTX 980 and as Accelerator, Intel Xeon Phi 31S1P. The main idea is compare each platform and evaluate which one fits best our needs.
In this work it intends give a solution to a problem hiperespectral images analysis field. This field is growing due to the analysis of hyperspectral images. Many algorithms have tried to solve this problem, this paper bring new algorithms to be accelerated under the OpenCl parallel programming paradigm. This paradigm give us the unified programming model to accelerate algorithms in heterogeneous systems. We can split the spectral unmixing in three stages. The first has the task of finding the number of pure pixels, called endmembers, pixels formed of a single spectral signature. Will be used the algorithm known as Geometry-based Estimation of number of endmember, GENE. The second stage is responsible for identifying endmember pixel and extract them along with all its bands. For this purpose we choose the algorithm called Simplex Growing Algorithm, SGA. In the last stage, making abundance maps of each endmember will be created. This step will be charged the algorithm known as, Sum-to-one Constrained Linear Spectral Unmixing, SCLSU. Three platforms have been used in this paper: As CPU, Intel Xeon E5-2695 v3, GPU, NVidia GeForce GTX 980 and as Accelerator, Intel Xeon Phi 31S1P. The main idea is compare each platform and evaluate which one fits best our needs.
Description
Trabajo de Fin de Grado en Ingeniería de Computadores (Universidad Complutense, Facultad de Informática, curso 2015/2016)