Publicación: Super-resolution : multi-frame registration and interpolation using optimal projections on functional spaces
Cargando...
Archivos
Fecha
2009-02-27
Autores
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
Atribución-NoComercial-SinDerivadas 4.0 Internacional
info:eu-repo/semantics/openAccess
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Educación a Distancia (España). Escuela Técnica Superior de Ingeniería Informática. Departamento de Informática y Automática
Resumen
El objetivo de las técnicas de súper-resolución (SR) es producir una imagen de alta resolución y alta calidad a partir de una serie de imágenes de baja resolución y escasa calidad de una misma escena. Las imágenes pueden provenir de muy variadas fuentes: un vídeo, una secuencia de fotos tomadas de una misma escena, imágenes médicas, de satélite, etc. En general, la SR resulta de utilidad cuando nos ha sido imposible tomar imágenes de una calidad adecuada debido a problemas de coste, limitaciones físicas del sistema u otras causas. La SR es uno de los problemas de tratamiento digital de imágenes que se encuentra más abierto hoy en día a pesar de los numerosos trabajos aparecidos desde el primer artículo que trataba este problema, el cual fue publicado en 1.984 (“Multi-frame image restoration and registration”, por Tsai y Huang). Entre las premisas que son necesarias para que la SR sea posible se encuentran que las imágenes presenten aliasing, es decir, que se encuentren muestreadas por debajo de la frecuencia de Nyquist, y que las imá- genes presenten movimiento entre ellas. De esta forma, cada imagen de baja resolución contiene distinta información y es posible extraer una imagen de mayor resolución de la secuencia. Los problemas que es necesario resolver para alcanzar la imagen de alta resolución final son: 1. El registro de las imágenes. Éste presenta una especial dificultad debido a que registramos imágenes con aliasing, por la que la información que poseen en común es menor que en otros problemas de procesamiento de imágenes. 2. Interpolación. Debemos pasar de una serie de muestras irregulares a un muestreo regular en rejilla que pueda ser fácilmente tratado algorítmicamente y visualizado. 3. Restauración. Aquí se incluye la reducción de ruido y la corrección de desenfoque debido a las lentes y al sensor. Estos problemas pueden resolverse de forma individual o bien conjuntamente, no pretendiendo sugerir con la lista anterior que el problema de la SR sea separable siempre de esta forma. Los métodos empleados en la SR han sido extremadamente variados: mé- todos en el dominio de la frecuencia, basados en proyección sobre conjuntos convexos, planteamiento de un sistema algebraico, aproximaciones polinómicas locales, basados en wavelets, etc. A pesar de esto, las técnicas de súper-resolución aún no han llegado a la madurez. En general, son bastante restrictivas en los modelos de movimiento permitidos, computacionalmente muy costosas y no siempre eficaces. Existe, por tanto, un amplio campo de posibles mejoras que permanece abierto. En esta tesis presentamos algoritmos novedosos en cada una de las etapas de la SR que hemos expuesto: 1. Un método de registro simultáneo de múltiples imágenes que establece relaciones entre las transformaciones de todas las imágenes para conseguir un registro más preciso. En los experimentos realizados demostramos un rendimiento superior con respecto a métodos clásicos de registro y otros métodos de registro para múltiples imágenes 2. Interpolación. Demostramos que el modelo implícitamente asumido en la mayoría de los métodos de SR lleva a la aparición de aliasing en las imágenes reconstruidas. Se propone un nuevo método que proyecta ortogonalmente sobre la base deseada las muestras de las imágenes de baja resolución, siguiendo la teoría clá- sica de muestreo. De esta forma, conseguimos que las imágenes reconstruidas no contengan artefactos debidos al aliasing, consiguiéndose así una mayor calidad tanto cuantitativamente como cualitativamente con respecto a otros métodos de SR. 3. Restauración. Uno de los mayores problemas en SR es la aparición de puntos no válidos (outliers) debidos a problemas del sensor, registro, etc. Para combatir estos problemas se recurre a métodos de SR robustos (es decir, con mecanismos de estimación robustos). Proponemos un filtro de la mediana sobre muestras irregulares como paso previo de eliminación de outliers antes de aplicar el algoritmo de SR. Hemos demostrado que este método consigue mejores resultados que otros algoritmos de SR robustos.
The objective of super-resolution (SR) techniques is to produce a high resolution and high quality image using as starting point a series of low resolution and low quality images from the same scene. The images can come from many sources: a video, a sequence of photographs from the same scene, medical images, satellite images, etc. In general, SR is useful when it has been impossible to obtain suitable images due to high cost problems, physical limitations of the system or other causes. Super-resolution is one of the imaging digital signal processing problems that is more open nowadays, besides the huge quantity of works that have appeared since the first paper that touched this problem, which was published in 1,984 (“Multi-frame image restoration and registration”, by Tsai and Huang). Among the premises that are necessary for SR to be possible is the need for aliasing to be present in the images, which means that the images must be sampled under the Nyquist frequency, with movement among themselves. In this way, each low resolution image contains different information and it is possible to extract a bigger resolution image from the sequence. The problems or stages that we need to solve to obtain the final high resolution image are: 1. The register of the images. It presents special difficulties due to the fact that we are registering aliased images, so the common information among them is less than in other image processing problems. 2. Interpolation. We must pass from a series of irregular samples to a regular sampling grid that can be addressed algorithmically and easily visualized. 3. Restoration. Here we include noise reduction and deblurring the distortions due to the lenses and the sensor. These problems can be solved individually or jointly: we do not intend to suggest with the previous list that the SR problem is always separable in this way. The methods employed in SR have been widely varied: methods in the frequency domain, based on projections on convex sets, posing an algebraic system, based on polynomial local approximations, using wavelets, etc. Besides of this, super-resolution techniques have not arrived yet to matureness. In general, they are quite restrictive in the allowed motion models, computationally very costly, and not always effective. Thus, there is an ample field of possible improvements that remains open. In this dissertation we present novel algorithms in each of the previously exposed SR stages: 1. A simultaneous registration method for multiple images that establishes relationships among the motion transformations of all the images to achieve a more accurate registration. In the presented experiments we demonstrate better performance with regard to classical registration methods and also to other multi-frame registration methods. 2. Interpolation. We demonstrate that the implicitly assumed data model in most SR methods leads to the appearance of aliasing artifacts in the reconstructed images. We propose a novel method that projects orthogonally on the desired basis the low resolution images samples, following classical sampling theory. In this way, we are able to remove aliasing artifacts in the reconstructed images, achieving quantitatively and visually better quality regarding to other SR methods. 3. Restoration. One of the biggest problems in SR is the appearance of outliers due to problems in the sensor, misregistration, etc. To address these problems, people resort to robust SR methods (that is, with robust estimation mechanisms). We propose a median filter for irregular samples to remove outliers as a previous step before applying the SR method of choice. We have demonstrated that this method is able to obtain better results than other robust SR algorithms.
The objective of super-resolution (SR) techniques is to produce a high resolution and high quality image using as starting point a series of low resolution and low quality images from the same scene. The images can come from many sources: a video, a sequence of photographs from the same scene, medical images, satellite images, etc. In general, SR is useful when it has been impossible to obtain suitable images due to high cost problems, physical limitations of the system or other causes. Super-resolution is one of the imaging digital signal processing problems that is more open nowadays, besides the huge quantity of works that have appeared since the first paper that touched this problem, which was published in 1,984 (“Multi-frame image restoration and registration”, by Tsai and Huang). Among the premises that are necessary for SR to be possible is the need for aliasing to be present in the images, which means that the images must be sampled under the Nyquist frequency, with movement among themselves. In this way, each low resolution image contains different information and it is possible to extract a bigger resolution image from the sequence. The problems or stages that we need to solve to obtain the final high resolution image are: 1. The register of the images. It presents special difficulties due to the fact that we are registering aliased images, so the common information among them is less than in other image processing problems. 2. Interpolation. We must pass from a series of irregular samples to a regular sampling grid that can be addressed algorithmically and easily visualized. 3. Restoration. Here we include noise reduction and deblurring the distortions due to the lenses and the sensor. These problems can be solved individually or jointly: we do not intend to suggest with the previous list that the SR problem is always separable in this way. The methods employed in SR have been widely varied: methods in the frequency domain, based on projections on convex sets, posing an algebraic system, based on polynomial local approximations, using wavelets, etc. Besides of this, super-resolution techniques have not arrived yet to matureness. In general, they are quite restrictive in the allowed motion models, computationally very costly, and not always effective. Thus, there is an ample field of possible improvements that remains open. In this dissertation we present novel algorithms in each of the previously exposed SR stages: 1. A simultaneous registration method for multiple images that establishes relationships among the motion transformations of all the images to achieve a more accurate registration. In the presented experiments we demonstrate better performance with regard to classical registration methods and also to other multi-frame registration methods. 2. Interpolation. We demonstrate that the implicitly assumed data model in most SR methods leads to the appearance of aliasing artifacts in the reconstructed images. We propose a novel method that projects orthogonally on the desired basis the low resolution images samples, following classical sampling theory. In this way, we are able to remove aliasing artifacts in the reconstructed images, achieving quantitatively and visually better quality regarding to other SR methods. 3. Restoration. One of the biggest problems in SR is the appearance of outliers due to problems in the sensor, misregistration, etc. To address these problems, people resort to robust SR methods (that is, with robust estimation mechanisms). We propose a median filter for irregular samples to remove outliers as a previous step before applying the SR method of choice. We have demonstrated that this method is able to obtain better results than other robust SR algorithms.
Descripción
Categorías UNESCO
Palabras clave
Citación
Centro
Facultades y escuelas::Escuela Internacional de Doctorado
Departamento
No procede