Visual Saliency Based On Natural Scene Statistics

Primates have a remarkable ability to detect salient objects in complex natural scenes in real time. They can quickly and accurately identify the targets despite the limited resources of the neural system. In recent years, how to draw the neural mechanisms of biological vision systems to design computer vision model receives more and more attention.Research of visual saliency is one of the hot issues in computer vision. Visual saliency is the perceptual quality that makes some items in visual scenes stands out from their immediate contexts. In other words,the areas which can cause obvious attention are salient. So far, there are lots of methods in visual saliency research, but there still exists limitations.In this paper, we propose a model called visual saliency based on natural scene statistics, which is derived from the study of biological vision. Different from the image features description in saliency measure, we developed a model of context-mediated probability distributions (PDs) in natural scene. In other words, saliency should be high when the visual variable appears a low probability to the context. The main steps include:first, we need to extract center-surround patches; second we compute the Independent Component (ICs) of the center-surround patches by Independent Component Analysis (ICA);third, we need to calculate conditional probability distribution; fourth, we obtain parameters of the PDs by generalized Gaussian distribution; last, we can get the saliency measure by the saliency formula.We studied the saliency in both static and dynamic natural scenes. To test the results of our saliency measure, we compared our model with some of existing algorithms, and used receiver operating characteristic (ROC) and Kullback-Leibler (KL) divergence to evaluate the efficiency. Experimental results show that our model has a good quality in predicting person's eye movements.