Publication: Implementación paralela del Automatic Target Detection and Classification Algorithm haciendo uso de la ortogonalización de Gram Schmidt para el análisis de imágenes hiperespectrales
Loading...
Official URL
Full text at PDC
Publication Date
2019
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
La observación remota de la Tierra ha sido siempre objeto de interés para el ser humano. A lo largo de los años los métodos empleados con ese fin han ido evolucionando hasta que, en la actualidad, el análisis de imágenes hiperespectrales constituye una línea de investigación muy activa, en especial para realizar la monitorización y el seguimiento de incendios o prevenir y hacer un seguimiento de desastres naturales, vertidos químicos u otros tipos de contaminación ambiental. Debido a la forma en que aparecen los materiales en el entorno natural, es muy habitual que cohabiten materiales diferentes en una misma porción de espacio, por pequeña que sea ésta. Ello hace que la gran mayoría de los píxeles analizados no siempre estén constituidos por la presencia de un único material, sino que estén formados por distintos materiales puros a nivel de subpíxel. Tradicionalmente, se utilizan para su análisis técnicas de desmezclado espectral que precisan de dos etapas complejas: la primera se basa en la extracción de firmas espectrales puras, también llamadas endmembers, y la segunda consiste en estimar el porcentaje de abundancia de dichos endmembers a nivel de subpíxel. Ambas etapas implican un alto coste computacional y esto supone un problema cuando se desea analizar imágenes hiperespectrales en tiempo real. Los algoritmos de clasificación y detección de objetivos (targets) tienen unos principios de funcionamiento muy similares a los algoritmos de detección de endmembers y por tanto, habitualmente se utilizan para este fin a pesar de su alto coste computacional. Las FPGAs (Field Programmable Gate Array) ofrecen suficiente rendimiento para realizar este proce samiento además de flexibilidad, pequeño tamaño y bajo consumo. Todo ello sumado a que pueden ser endurecidas para su uso en el espacio hacen de ésta una opción muy viable para el objetivo que se persigue. En este trabajo de fin de grado se lleva a cabo la implementación en FPGA del algoritmo de detección y clasificación de objetivos conocido como ATDCA-GS (Automatic Target Detection and Classification Algorithm - Gram Schmidt), que utiliza el concepto de proyección ortogonal de un subespacio, haciendo uso de la ortogonalización de Gram Schmidt para simplificar operaciones complejas. Para la consecución de dicho TFG, se ha realizado la implementación del algoritmo en el lenguaje de descripción hardware VHDL (Very High Speed Integrated Circuit Hardware Description Language) y además, se ha analizado y optimizado otra implementación previa bajo el paradigma de programación paralela OpenCL. Posteriormente, se han comparado ambas implementaciones optimizadas en términos de precisión y rendimiento sobre plataformas de hardware reconfigurable tipo FPGAs.
The remote observation of the Earth has always been an object of interest for the human being. Over the years, the methods used for this purpose have evolved until, at present, the analysis of hyperspectral images constitutes a very active line of research, especially to monitor fires or prevent and monitoring natural disasters, chemical discharges or other types of environmental pollution. Due to the way in which materials appear in the natural environment, it is very common to cohabit different materials in the same portion of space, however small it may be. This means that the vast majority of pixels analyzed are not always constituted by the presence of a single material, but are formed by different pure materials at the sub-pixel level. Traditionally, spectral demixing techniques are used for their analysis, which require two complex stages: the first is based on the extraction of pure spectral signatures, also called endmembers, and the second consists of estimating the percentage of abundance of said endmembers at the level of sub-pixel. Both stages involve a high computational cost and this is a problem when you want to analyze hyperspectral images in real time. The algorithms of classification and detection of targets have some operating principles very similar to the detection algorithms of endmembers and therefore, they are usually used for this purpose despite their high computational cost. FPGAs (Field Programmable Gate Array) offer sufficient performance to make this processing as well as flexibility, small size and low consumption. All this together with the fact that they can be hardened for use in the space make this a very viable option for the objective pursued. In this End-of-Degree Project, the FPGA implementation of the Automatic Target Detection and Classification Algorithm - Gram Schmidt known as ATDCA-GS, which uses the concept of orthogonal projection of a subspace, is carried out, using the orthogonalization of Gram Schmidt to simplify complex operations. To achieve this proyect, the algorithm has been implemented in the hardware description language VHDL (Very High Speed Integrated Circuit Hardware Description Language) and in addition, another previous implementation under the OpenCL parallel programming paradigm was analyzed and optimized. Subsequently, both implementations optimized in terms of accuracy and performance have been compared on reconfigurable hardware platforms such as FPGAs.
The remote observation of the Earth has always been an object of interest for the human being. Over the years, the methods used for this purpose have evolved until, at present, the analysis of hyperspectral images constitutes a very active line of research, especially to monitor fires or prevent and monitoring natural disasters, chemical discharges or other types of environmental pollution. Due to the way in which materials appear in the natural environment, it is very common to cohabit different materials in the same portion of space, however small it may be. This means that the vast majority of pixels analyzed are not always constituted by the presence of a single material, but are formed by different pure materials at the sub-pixel level. Traditionally, spectral demixing techniques are used for their analysis, which require two complex stages: the first is based on the extraction of pure spectral signatures, also called endmembers, and the second consists of estimating the percentage of abundance of said endmembers at the level of sub-pixel. Both stages involve a high computational cost and this is a problem when you want to analyze hyperspectral images in real time. The algorithms of classification and detection of targets have some operating principles very similar to the detection algorithms of endmembers and therefore, they are usually used for this purpose despite their high computational cost. FPGAs (Field Programmable Gate Array) offer sufficient performance to make this processing as well as flexibility, small size and low consumption. All this together with the fact that they can be hardened for use in the space make this a very viable option for the objective pursued. In this End-of-Degree Project, the FPGA implementation of the Automatic Target Detection and Classification Algorithm - Gram Schmidt known as ATDCA-GS, which uses the concept of orthogonal projection of a subspace, is carried out, using the orthogonalization of Gram Schmidt to simplify complex operations. To achieve this proyect, the algorithm has been implemented in the hardware description language VHDL (Very High Speed Integrated Circuit Hardware Description Language) and in addition, another previous implementation under the OpenCL parallel programming paradigm was analyzed and optimized. Subsequently, both implementations optimized in terms of accuracy and performance have been compared on reconfigurable hardware platforms such as FPGAs.
Description
Trabajo de Fin de Grado, Universidad Complutense, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática. Curso 2018/2019