Nonlinear data fusion in saliency-based visual attention

Abstract

Primates use saliency-based visual attention to detect conspicuous objects in cluttered visual environments. Some new strategies of combining feature maps to form a saliency map are addressed in this paper. Traditional methods of making saliency map are to linearly combine feature maps extracted from early visual system. Here we have proposed some modifications in saliency model with three different data fusion schemes: weighted linear combination of feature maps, multiplicative saliency map, and harmonic mean of feature maps. Experiments are based on a 32 images dataset of emergency triangle in natural environments. Comparison with the basic saliency model has also been provided. Results suggest that nonlinear combination of feature activities could perform a more accurate detection, and speeds up the process of finding a desired object in the scene.