Visual saliency detection via integrating bottom-up and top-down information

Abstract

Selective attention is a process that enables biological and artificial systems to remove the redundant information and highlight the valuable regions in an image. The relevant information is determined by task driven (Top-Down (TD)) or task-independent (Bottom-Up (BU)) factors. In this paper, we present a new computational visual saliency model which uses the combination of BU and TD mechanism for extracting the relevant regions of images with man-made objects. The prior knowledge about man-made objects is the compactness and higher values of different orientations. So, by using maximum and minimum moments of phase congruency covariance and different orientations from Gabor filters, we obtain different feature maps from two mentioned attention mechanisms. Finally these maps are linearly combined which their coefficients are obtained by using the entropy of each feature map. Three region-based databases were used to examine the performance of the proposed method. The experimental results demonstrated the efficiency and effectiveness of this new visual saliency model.