Visual Attention Model And Its Application On Scene Categorization

Visual attention is known as an important feature of human visual system. Human can refuse useless information, and select the useful information of complex scene by visual attention mechanism. Visual attention can guide human attention on salient region which is different from surround environment, can also guide human attention on target that human are looking for. Because computational model of visual attention has an important application on object recognition, object tracking, image analysis and understanding and so on, researching on it becomes the focus in computer vision.In this paper, a model called multiscale visual saliency using natural statistics is proposed, which is derived from scale space theory and natural statistics. The saliency measure is based on natural image statistics, rather than based on a single test image. The main steps include:first, extract multiscale patches from natural scene images dataset; second, compute the Independent Component(ICs) of patches by Independent Component Analysis(ICA), and probability distributions(PDs) of natural scenes are represented by independent components of natural scenes; third, estimate the parameters of the PDs in natural scenes by Generalized Gaussian Distribution(GGD) fitting; last, compute saliency maps at all scales separately, and get combined saliency map by getting highest saliency at each position. Furthermore, we access how well our model of visual saliency in both static and dynamic natural scenes predicts human performance.Visual saliency is applied to scene categorization. A scene is represented by the patches extracted from the salient regions. Then each patch is assigned to a visual word. Topic models in document analysis are used for scene classification. We test two classifiers, SVM and Bayes classifier. Simulation results demonstrate that the categorization accuracy with SVM classifier is better than with Bayes classifier.