A saliency-based bottom-up visual attention model for dynamicscenes analysis

This work proposes a model of visual bottom-up attention for dynamic scene analysis. Our work addsmotion saliency calculations to a neural network model withrealistic temporal dynamics [(e.g., building motion salienceon top of De Brecht and Saiki Neural Networks 19:1467–1474, (2006)]. The resulting...

ver descrição completa

Detalhes bibliográficos
Autores: Ramírez Moreno, David Fernando, Schwartz, Odelia, Ramirez Villegas, Juan Felipe
Formato: artículo
Estado:Versión publicada
Fecha de publicación:2013
País:Colombia
Recursos:Universidad Autónoma de Occidente
Repositorio:RED: Repositorio Educativo Digital UAO
Idioma:inglés
OAI Identifier:oai:red.uao.edu.co:10614/11584
Acesso em linha:http://red.uao.edu.co//handle/10614/11584
Access Level:acceso abierto
Palavra-chave:Redes neuronales (Computadores)
Neural networks (Computer science)
Visual attention
Saliency map
Motion saliency
Neural network
Synaptic depression
Neural latency
Asymmetry phenomenon
Lyapunov stabilit
Descrição
Resumo:This work proposes a model of visual bottom-up attention for dynamic scene analysis. Our work addsmotion saliency calculations to a neural network model withrealistic temporal dynamics [(e.g., building motion salienceon top of De Brecht and Saiki Neural Networks 19:1467–1474, (2006)]. The resulting network elicits strong transientresponses to moving objects and reaches stability withina biologically plausible time interval. The responses arestatistically different comparing between earlier and latermotion neural activity; and between moving and non-movingobjects. We demonstrate the network on a number of syn-thetic and real dynamical movie examples. We show thatthe model captures the motion saliency asymmetry phenom-enon. In addition, the motion salience computation enablessudden-onset moving objects that are less salient in the staticscene to rise above others. Finally, we include strong consid-eration for the neural latencies, the Lyapunov stability, andthe neural properties being reproduced by the model