Divergence Filter for Saliency

Detection of regions with high visual attention from image has various applications including advertising design where ads are often associated with relevant semantic visual information. The salient regions in the image/video have to be identified in a consistent way, even if original objects or background are texture scene. This is achieved by solving combinatorial problem of down-sampling that searches for the optimal texture region map. The complexity of this solution makes it impractical. The problem becomes easy by a new approach for saliency detection. It is based on the spatial attention model that evaluates divergence of a given local region from its surrounding where objects and background can be texture scene. Our proposed solution is based on an adaptive version of the bilateral filter that searches for the divergence of a pixel with its local neighbors. The contributions of this work are new divergence estimation function which reduces potential global search into a simple local filter, and efficient convex-hull algorithm for creating saliency map. Experimental results show that the solution can deal with texture during analysis of visual attention, and saliency detection's performance is improved.