Publication: Implementación sobre FPGA de un algoritmo de compresión de imágenes hiperespectrales basado en JPEG2000
Loading...
Official URL
Full text at PDC
Publication Date
2018
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Las imágenes hiperespectrales son uno de los métodos de análisis disponibles hoy en día para estudios no intrusivos, a distancia, de cualquier elemento. Capturando cientos de longitudes de onda en cada píxel, la información que proporcionan, además de útil y precisa, es pesada. Una sola imagen puede superar el Giga Byte, por lo que la compresión es más una obligación que una opción. Si queremos reducir en una fracción importante el tamaño, debemos ceñirnos a los algoritmos con pérdida. En los últimos años han ido evolucionando, con las técnicas más prometedoras utilizando híbridos entre algoritmos de compresión para imágenes tradicionales, y técnicas que incorporan una decorrelación espectral en cada píxel mediante reductores dimensionales. En este trabajo de fin de máster se ha partido de esa idea, diseñando un compresor con múltiples reductores dimensionales, utilizando como núcleo de la compresión las ideas del estándar JPEG2000, que han sido ampliadas con numerosas opciones. Los resultados muestran niveles de compresión por encima de los obtenidos anteriormente, con la elección de parámetros jugando un papel fundamental en la calidad de las imágenes comprimidas. A partir de ese desarrollo, se realizó un análisis de tiempos del algoritmo, detectando las partes más lentas. Mediante técnicas de submuestreo en la reducción dimensional, los tiempos se mejoraron sin afectar a la precisión. Se vio además la posibilidad de conseguir mejoras adicionales incorporando una FPGA como coprocesador para la codificación de JPEG2000. La implementación VHDL ha dado excelentes resultados, y gracias a sus características es arbitrariamente paralelizable, haciendo que la codificación sea prácticamente instantánea en las FPGA con más capacidad del mercado. Juntando todas estas ideas, se ha conseguido un compresor híbrido capaz de reducir los tiempos de cálculo de varios minutos a meros segundos, posibilitando la compresión con pérdida en tiempo real, a la vez que se mantiene un alto rendimiento de distorsión frente a ratio de compresión.
Hyperspectral images are one of the latest methods for non-intrusive analysis, allowing distant studies of virtually any subject. Capturing hundreds of different wavelengths in every pixel they provide valuable and precise, yet heavyweight, information. An image alone can surpass a GigaByte, so compression is no longer optional. If our goal is to significantly reduce image size, we must focus on lossy algorithms. They have been evolving in the past years, with the most promising techniques being a hybrid between traditional image compression algorithms and spectral decorrelation at the pixel level, using dimensionality reduction. This master’s thesis is based on that idea. A compressor with multiple dimensionality reduction algorithms is designed, with the ideas of the JPEG2000 standard at its core, and is extended with new options. Results show compression ratios higher than those previously obtained, with the choice of parameters playing a crucial role in compression quality. Based on those findings, the different steps of the algorithm were time profiled to find the slowest bottlenecks. Execution times were greatly improved, without a hit in accuracy, using subsampling techniques in the training steps of the dimensionality reduction algorithms. It was also found that the coding step of the JPEG2000 algorithm could be significantly sped up on custom hardware. To address this, an FPGA implementation was designed using VHDL, which greatly improved execution times thanks to the parallelizable nature of the algorithm. Putting all these ideas together, a hybrid compressor is presented which reduces compression time from minutes to seconds, allowing real-time lossy compression with a great rate-distortion performance.
Hyperspectral images are one of the latest methods for non-intrusive analysis, allowing distant studies of virtually any subject. Capturing hundreds of different wavelengths in every pixel they provide valuable and precise, yet heavyweight, information. An image alone can surpass a GigaByte, so compression is no longer optional. If our goal is to significantly reduce image size, we must focus on lossy algorithms. They have been evolving in the past years, with the most promising techniques being a hybrid between traditional image compression algorithms and spectral decorrelation at the pixel level, using dimensionality reduction. This master’s thesis is based on that idea. A compressor with multiple dimensionality reduction algorithms is designed, with the ideas of the JPEG2000 standard at its core, and is extended with new options. Results show compression ratios higher than those previously obtained, with the choice of parameters playing a crucial role in compression quality. Based on those findings, the different steps of the algorithm were time profiled to find the slowest bottlenecks. Execution times were greatly improved, without a hit in accuracy, using subsampling techniques in the training steps of the dimensionality reduction algorithms. It was also found that the coding step of the JPEG2000 algorithm could be significantly sped up on custom hardware. To address this, an FPGA implementation was designed using VHDL, which greatly improved execution times thanks to the parallelizable nature of the algorithm. Putting all these ideas together, a hybrid compressor is presented which reduces compression time from minutes to seconds, allowing real-time lossy compression with a great rate-distortion performance.
Description
Máster en Ingeniería Informática, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática, curso 2017-2018