Computational Modeling of Saliency from Image Histogram

Abstract

Computational model of image saliency plays an important role for vision tasks such as visual search and attention modeling. We developed a computational model that captures both shape and color image saliency based on histograms. The designed model has been evaluated over a set of fixation maps of 120 natural images that are recorded from 20 subjects for the purpose of saliency computation within visual cortex [Neil D. B. Bruce, et al, 2009]. We present four key observations. First, saliency in both image color and image shape can be efficiently computed by histograms that encode both local and overall distributions of image values in different color channels. Second, due to the use of image histogram the designed saliency model is much more tolerant to the variation of image scales compared with other common saliency models. Third, the designed saliency model is much more tolerant to image edges whereas other common saliency models are often biased towards image edges especially when the saliency is computed on a large image scale. Last but not least, the designed saliency model shows that compared with image brightness, image color contributes much more to the overall image saliency for natural scene images.