Evolving sensor environments with visual attention: An experimental exploration

Abstract

The research community is experiencing nowadays a significant growth in the amount of sensor data made available to several practical applications, particularly those dealing with visual information. The availability of large datasets poses critical challenges for the selection of only relevant features to allow their timely use and interpretation. The recent years marked an increasing interest in algorithms inspired from biological human vision as an alternative source of ideas for the development of computational resources to deal with large datasets. In particular, computational models of visual attention have been shown to significantly improve the speed of scene understanding and object recognition by attending only the regions of interest and distributing the resources where they are required. This paper explores the use and gauges the performance of visual attention mechanisms for identifying an optimal feature set that ensures the identification and classification of objects in images, in two different scenarios. The first scenario addresses the issue of the identification of different categories of vehicles from multiple viewpoints in a controlled environment, with a relatively known background. The other scenario explores the capabilities of an improved visual attention model for the identification of buildings in satellite imaging, characterized by large variations in content and characteristics and by a cluttered background.