Multi-Level Paralelism in Image Identification
Juan Alvaro Fernandez, R. Guerrera, Natalia Miranda, Fabiana Piccoli · Mecánica Computacional (Asociación Argentina de Mecánica Computacional) · 2009
Images represent complex visual information. From a human point of view, image information extraction is merely a global process: starting with usually short, local and bad quality information, our brain can identify basic compositing elements and their relationship, elaborate a coherent sensorial experience and finally get the required information. From a computational point of view, images clearly represent non structured information. Getting information from them involves to simulate human visual perception. At early stages, a human visual system is involved in identifying objects (the “ what ” system) and in locating them (the “where” system). The two systems can be simulated by looking for visual cues such as color , shape , depth , and motion. The combined cues lead to a unique image visual content representation. The most common image representation is a feature vector, where each vector component represents an image feature. At this work we proposed a parallel feature vector construction following the precepts stated by the “what ” and “ where ” system and by using high performance computing. The problem enables the application of multiple levels of parallelism and different paradigms combination. Finally some p reliminary results are given.