| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.14 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Ao longo dos anos, as informações provenientes de conteúdos multimédia têm sido
cada vez mais usadas, sendo um fator muito importante na área da comunicação e da
interação. De forma a ajudar na organização, descrição e armazenamento dessas informações,
foram criados mecanismos de indexação e de acesso baseados nos conteúdos visuais.
Um dos primeiros passos identificados para estes mecanismos é a análise estrutural de
vídeos, onde é efetuada a extração de cenas, detetando os limites das mesmas. Sendo esta
temática um objeto de investigação há já algum tempo, existe uma variedade de técnicas
propostas e avaliadas que vão desde a simples comparação de frames adjacentes até a
métodos que envolvem aprendizagem automática. Porém, estes métodos são computacionalmente
pesados, apresentando tempos de processamento perto da metade da duração
dos vídeos analisados.
Nesta dissertação propomos uma arquitetura distribuídaMulti-CPU eMulti-GPU que
faz a extração de features presentes em frames de vídeo, efectuando a deteção dos cortes
de cena. O nosso objetivo é tirar partido do poder computacional presente num cluster
híbrido de GPUs e CPUs, deste modo fazemos uma distribuição do processamento de
vídeos pelos vários nós do cluster.
Over the years, the usage of information provided by multimedia resources has been increasing: an essential factor in communication and interaction. To organize, describe and store that information, mechanisms of indexing and retrieval based on visual content have been created. One of the identified steps of these mechanisms is a video structural analysis. When a shot extraction is performed we can detect shot boundaries. This topic has been subject to investigation multiple times and a large variety of techniques have been proposed and evaluated. These techniques can vary in complexity as a very simple comparison of adjacent frames to Machine Learning. However, these methods are computationally intensive leading to processing times close to 50% of the examined videos. In this dissertation, we propose a distributed CPU-GPU architecture that makes the extraction of features present in video frames to make shot boundary detections. Our goal is to take advantage of the computational power of CPU and GPU devices present in a cluster. To that end, we did a distribution of the video processing to the nodes of the cluster.
Over the years, the usage of information provided by multimedia resources has been increasing: an essential factor in communication and interaction. To organize, describe and store that information, mechanisms of indexing and retrieval based on visual content have been created. One of the identified steps of these mechanisms is a video structural analysis. When a shot extraction is performed we can detect shot boundaries. This topic has been subject to investigation multiple times and a large variety of techniques have been proposed and evaluated. These techniques can vary in complexity as a very simple comparison of adjacent frames to Machine Learning. However, these methods are computationally intensive leading to processing times close to 50% of the examined videos. In this dissertation, we propose a distributed CPU-GPU architecture that makes the extraction of features present in video frames to make shot boundary detections. Our goal is to take advantage of the computational power of CPU and GPU devices present in a cluster. To that end, we did a distribution of the video processing to the nodes of the cluster.
Descrição
Palavras-chave
Análise de Vídeo Deteção de Cortes de Cena OpenCL Computação Distribuída Computação Heterogénea
